perf: Bound Whisper audio decoding memory

## What
Decode, downmix, resample, and feed Whisper audio in bounded chunks instead of holding the full decoded recording in memory.

## Why
- `.planning/codebase/CONCERNS.md` notes that transcription reads the whole WAV before resampling and chunking.
- `src-tauri/src/whisper/mod.rs` currently builds full sample vectors before inference.
- Long meetings can allocate a full PCM buffer plus normalized `f32` audio before transcription starts.

## How
- Stream WAV decoding through a bounded processing pipeline.
- Downmix and resample incrementally, then pass fixed-size chunks to Whisper.
- Consider an intermediate normalized temp file if direct streaming is too complex.
- Add a regression test for empty and very short inputs plus a memory-oriented smoke test for longer fixtures.