A correctness release, with Native AOT as the theme. NAudio's remaining [StructLayout] interop types — WAVEHDR, ACMSTREAMHEADER and WaveFormat itself — were passed to native code by value, which CoreCLR pins in place but the AOT marshaller copies into a per-call temporary. Anything the driver or codec wrote back was silently discarded, so WaveIn/WaveOut, ACM conversion (and with it Mp3FileReader and AudioFileReader) and every WASAPI, Media Foundation and DMO format negotiation were broken in an AOT-published app while building and running cleanly. Those paths now marshal by hand into stable unmanaged blocks, and NAudio.WinMM joins the AOT-compatible package set.

The two breaking changes below both fall out of that work. Also fixed are two 3.0.0 regressions — Mp3FileReaderBase seeking on files with a Xing/Info header, and WdlResamplingSampleProvider losing samples when its source under-feeds — and a pair of infinite loops in the WAV parser on corrupt files.

• Breaking: WaveFormat and its subclasses no longer carry [StructLayout], so Marshal.SizeOf and Marshal.PtrToStructure on them now throw ArgumentException instead of silently returning wrong data under Native AOT (Marshal.SizeOf<WaveFormatExtensible>() returned 22 rather than 40, and a decoded format kept its constructor defaults). NAudio itself stopped marshalling these types in 3.1.0; use ToWaveFormatExBytes, MarshalToPtr or MarshalFromPtr to cross a native boundary (#1432)
• Breaking: AudioClient.IsFormatSupported and WasapiPlayer.IsFormatSupported now return the closest-match format as out WaveFormat rather than out WaveFormatExtensible. WASAPI documents this format as either a WAVEFORMATEX or a WAVEFORMATEXTENSIBLE, and the old signature could only represent the latter — a plain format was previously decoded as a WaveFormatExtensible whose base fields were valid but whose extensible fields were meaningless. Callers needing the extensible fields should pattern-match (#1425)
• Fixed WaveIn/WaveOut failing under Native AOT with WaveHeaderUnprepared ("The wave header was not prepared"). The WAVEHDR was a [StructLayout] class passed to winmm by value, which CoreCLR pins in place but NativeAOT copies into a per-call temporary, so the driver's WHDR_PREPARED/WHDR_DONE/dwBytesRecorded writes were discarded. It is now a struct in a stable unmanaged block (#1425)
• Fixed ACM conversion failing under Native AOT — WaveFormatConversionStream, AcmStream and so Mp3FileReader and AudioFileReader. ACMSTREAMHEADER was a [StructLayout] class passed to msacm32 by value, and the codec keeps private state in the header's reserved fields between acmStreamPrepareHeader, acmStreamConvert and acmStreamUnprepareHeader; the AOT marshaller's per-call copy round-trips only the declared fields, so that state was lost and conversion failed. It is now a struct in a stable unmanaged block, as WAVEHDR already is (#1425)
• Fixed WaveFormat.MarshalToPtr and WaveFormat.MarshalFromPtr corrupting any WaveFormat subclass under Native AOT — the AOT marshaller drops the inherited WAVEFORMATEX fields of a class hierarchy, so a WaveFormatExtensible was written with its SubFormat GUID over the sample rate. Every remaining site that sized or wrote a WaveFormat with Marshal.SizeOf/StructureToPtr was converted too: WASAPI (AudioClient), Media Foundation (MFInitMediaTypeFromWaveFormatEx, reachable from MediaFoundationResampler/MediaFoundationEncoder), DMO (DmoMediaType.SetWaveFormat, MediaObject) and ACM (AcmStream, AcmDriver) (#1425)
• WaveFormat.MarshalToPtr now always allocates at least the 18 + cbSize bytes it advertises. A WaveFormat subclass that declares extraSize but doesn't write it in Serialize previously produced a block shorter than its own cbSize, which a native consumer would read past (#1425)
• Added WaveFormat.ToWaveFormatExBytes(), which renders a WaveFormat as a native WAVEFORMATEX byte array. MarshalToPtr is now a thin wrapper over it, for callers who need an unmanaged block rather than a buffer they already own (#1425)
• Added a Serialize override to Mp3WaveFormat, which previously advertised cbSize = 12 but wrote none of its 12 MPEGLAYER3WAVEFORMAT extra bytes (#1425)
• WaveFormatExtraData now sizes its buffer from cbSize instead of holding a fixed 100-byte array, so a fmt chunk declaring more than 100 bytes of extra data keeps it rather than having all of it discarded (#482). The fixed size existed only because [MarshalAs(ByValArray, SizeConst = 100)] needed a compile-time length. ExtraData is now exactly ExtraSize bytes long rather than always 100 (#1432)
• NAudio.WinMM is now marked IsAotCompatible, and the AOT smoke test covers the winmm WAVEHDR/WAVEFORMATEX paths (#1425)
• Fixed Mp3FileReaderBase seeking silently restarting playback from the beginning of the file on MP3s with a Xing/Info header — the lazy frame index was gated on IsLengthExact, which such a header sets without any frame having been scanned. Also fixed seeks landing on the wrong frame when the target fell exactly on a frame boundary, and Xing/Info header frames being indexed as audio (shifting every seek in those files ~26 ms early). A 3.0.0 regression (#1419)
• Fixed WdlResamplingSampleProvider losing samples, and eventually returning 0 permanently, when asked for more output than the source could supply — a 3.0.0 regression that broke the common pattern of reading generously from a BufferedWaveProvider-backed capture chain. WdlResampler.ResampleOut also no longer drifts in input-driven (feed) mode when handed fewer samples than ResamplePrepare requested (#1412)
• Fixed two infinite loops in the WAV parser triggered by corrupt files declaring a negative chunk size. An RF64 ds64 chunk with a negative dataChunkLength, or a LIST/adtl sub-chunk with a negative size, made the chunk walk advance by zero bytes per iteration and spin at 100% CPU without throwing. Negative and oversized sizes are now rejected where they are read, and an undersized ds64 chunk throws FormatException rather than ArgumentOutOfRangeException (#1428)
• Fixed the NAudioConsoleTest WASAPI exclusive-mode quick scan reporting no supported formats at all on devices that support plenty — it built its probe formats with an empty WAVEFORMATEXTENSIBLE channel mask, which drivers such as Realtek reject outright. It now probes with the canonical layout for the channel count, the same mask the library's own format adaptation uses (#1431)