feat: add MP3 streaming encoder (item 8's MP3 sub-item)

[cloud-hai-vo]

Summary

• Implements the MP3 sub-item of docs/streaming-plan.md item 8: Mp3StreamingEncoder is the third IStreamingEncoder implementation, reusing Mp3FrameEncoder's own encode internals unchanged — only how each encoded frame is delivered differs (a callback emitting raw, self-delimiting MPEG frames instead of writing bytes to a file).
• Writes no dedicated "Info tag" (LAME-tag) header frame and no ID3v2 tag at all — both are container/file-level conventions with no equivalent for a raw, open-ended frame stream, the same reasoning that already excluded ADTS framing from AacStreamingEncoder's own raw access units (PR feat: add push-style IStreamingEncoder + AAC-LC streaming session #110) and Ogg/pre_skip from OpusStreamingEncoder's own raw packets (PR feat: add Opus streaming encoder (item 8's Opus sub-item) #118). Mp3Encoder.EncoderDelaySamples (the existing, already empirically-measured 1680) remains the right value for a muxer here to use directly.
• Unlike OpusStreamingEncoder, this encoder reports each frame's own real sample count rather than always the nominal 1152 — tracking it costs nothing extra here, the same way it doesn't for AacStreamingEncoder, since Mp3EncoderSession.Finish's own existing design already computes the real remaining count before padding it.
• Complete()'s own mandatory flush frame (draining the one-granule MDCT history — the same real, previously-found-and-fixed requirement Mp3EncoderSession.Finish itself documents) is purely synthetic and reports a real sample count of 0.
• FLAC streaming sessions (item 8's remaining sub-item) remain not started.
• Docs synced: docs/streaming-plan.md (MP3 sub-item marked done), CLAUDE.md, README.md, llms.txt, llms-full.txt.

Test plan

• New Mp3StreamingEncoderTest: construction validation (null callback, unsupported channels/sample rate/bitrate), SamplesPerFrame == 1152, Complete() with zero samples emits nothing, idempotent Complete()/safe double Dispose(), Write after Complete() throws, monotonic frame indices, exact per-frame SampleCount verification (real count for the padded tail frame, 0 for the synthetic flush frame — including the "exactly one full frame, nothing pending" edge case), real round-trip SNR for mono and stereo (concatenating raw frames into a file, decoding via the project's own Mp3Decoder/NLayer), and a regression test mirroring Mp3EncoderSnrTest's own exact-tail-preservation check, confirming this restructured callback-based version of that fix still holds.
• Full existing suite (dotnet test, net10.0 locally): 1703 passed, 1 pre-existing unrelated skip, 0 failed.
• Full solution build (all target frameworks + AotSmokeTest): 0 warnings/errors.

🤖 Generated with Claude Code