feat: add local whisper-cli transcription backend

[Zaargh]

Summary

• Adds local whisper-cli (whisper.cpp) as a third transcription backend, preferred over Groq/OpenAI API when available
• Auto-detects binary via PATH and GGML models in ~/.cache/whisper/ or brew cellar
• Extracts audio as 16kHz mono WAV (native format for whisper.cpp) instead of mp3
• New CLI flag --whisper local to force local backend
• Config: WHISPER_MODEL env var for explicit model path, WHISPER_LOCAL=false to disable
• Updates setup.py preflight, session hook, and installer to detect and report local whisper

Motivation

Users with whisper-cpp installed locally can transcribe without API keys, network access, or cost. Local transcription is also more private since audio never leaves the machine.

[Zaargh]

Rebased onto main — conflicts in CHANGELOG.md (kept the new 0.1.3 release block alongside the Unreleased entry) and scripts/whisper.py (kept out_path.resolve() from the subprocess hardening in #2, dropped the redundant -b:a 64k since codec args are now branched on wav vs mp3).

PR is MERGEABLE / CLEAN again. Friendly bump in case it slipped past — happy to address any review feedback.