fix(cli): time out stalled model downloads

[miguel-heygen]

Root cause

The shared model downloader had no inactivity timeout. If a server opened a successful response but stopped delivering bytes, pipeline() waited forever and left a zero-byte .tmp file. This surfaced while Kokoro downloaded kokoro-v1.0.onnx.

Fix

• apply a 30-second socket inactivity timeout to each download request
• destroy stalled requests with an actionable timeout error
• remove partial files consistently on timeout, request error, pipeline failure, and non-200 response
• drain redirect/error response bodies

The timeout measures inactivity, not total download duration, so slow downloads continue while bytes are arriving.

Regression coverage

Adds a red-first test with a 200 response that never delivers data. It verifies prompt rejection and cleanup of both the partial and destination files.

Validation

• full repository build
• CLI suite: 130 files, 1,647 tests passed
• CLI typecheck
• focused downloader/TTS/Whisper tests
• changed-file lint, format, fallow, and pre-commit hooks

Reviewer action

Confirm the 30-second inactivity window is appropriate for model hosts; it does not cap total transfer duration.