CLI: give each input its own output name; report audio errors clearly

[CodeWithBehnam]

What does this PR do?

Closes #7.

• Output names: the output name was kept from the first input, so vayu a.mp3 b.mp3 wrote both transcripts to a.*. Each input now uses its own stem (content for stdin). Two cases are rejected before any work starts: --output-name with several inputs, and inputs that share a stem (e.g. day1/talk.mp3 day2/talk.mp3).
• stdin: stdin is now read inside the per-file error handling. Errors report the input path, not the decoded array.
• load_audio errors: a missing file raises FileNotFoundError. A missing ffmpeg, or ffmpeg failing to decode, raises the new AudioLoadError (a RuntimeError, so existing except RuntimeError still works). The CLI's old FileNotFoundError/CalledProcessError handlers never fired, and a missing ffmpeg was reported as "File not found: ".
• Early check: transcribe() checks the audio path exists before loading the model, so a typo no longer triggers a model download first.
• AudioLoadError is exported from the package.

Behaviour change: load_audio used to raise ValueError for a missing file; it now raises FileNotFoundError.

How was this tested?

• Tested with audio file(s)
• Ran existing tests (pytest): 44 passed, 1 skipped (the ffmpeg-decoding test skips without ffmpeg; it passed locally with an ffmpeg binary)
• Tested CLI (vayu audio.mp3): tests/test_cli.py drives cli.main() with stubbed transcription

tests/test_cli.py has 11 tests. Eight fail on main.

claude-review will fail as on #12 (the repository's CLAUDE_CODE_OAUTH_TOKEN secret, see this comment).

🤖 Generated with Claude Code

https://claude.ai/code/session_01TMXYMqgLAykqRApmbRfpTA

---

Generated by Claude Code