fix(transcription): scale local transcription timeouts with audio duration

[xAlcahest]

Summary

Local transcription dies after exactly 5 minutes on long recordings: the local single-request paths hardcode a 300000 ms request timeout (whisperServer.js POST /inference, parakeetWsServer.js transcribe, llama post-processing, diarization), and the wall is on inference time, so a long file on a slow CPU with a big model gets killed mid-transcription. The new src/helpers/transcriptionTimeout.js scales the timeout with the estimated audio duration (max(300000, seconds * 10000), with a 24-hour ceiling to stay under the 32-bit setTimeout limit) and uses a flat 60-minute cap where duration is not cheaply available, so short dictations keep exactly today's behavior and a genuinely hung server still fails eventually.

Fixes #976

Changes

• src/helpers/transcriptionTimeout.js: new shared module with the floor, per-second, flat-cap and ceiling constants plus computeTranscriptionTimeoutMs()
• src/helpers/whisperServer.js: /inference request timeout scales with the converted WAV buffer duration
• src/helpers/parakeetWsServer.js: per-call timeout computed from the sample buffer instead of the flat 5-minute constant
• src/helpers/llamaServer.js: post-processing timeout uses the 60-minute flat cap since no audio duration exists there
• src/helpers/diarization.js: per-run timeout derived from the WAV file size, falling back to the flat cap
• test/helpers/transcriptionTimeout.test.js: floor, scaling, ceiling and invalid-input coverage