Add offline plan_segments() seam for batch transcription

[sumgup0]

Exposes TransClip's silence-cut policy as a pure, offline plan_segments(pcm16, *, sample_rate, max_segment_s) so a downstream meeting recorder can reuse it. Returns contiguous, non-empty PlannedSegment intervals that partition the whole input; cuts at the latest qualifying silence per window (reusing the live commit-cut logic) and forces a cut at the quietest frame in the window's final FORCED_CUT_WINDOW_S when far-end speech never pauses. No segment exceeds max_segment_s.

Dictation path unchanged. Commit 1 is a pure, behavior-preserving extraction: _frame_dbfs/_silence_threshold are pulled out of _find_commit_cut with byte-identical arithmetic; its signature and its live caller are untouched. Commit 2 adds the offline planner and its tests. The planner lives in asr_incremental.py (not a new module) so it can reuse those private commit-cut helpers.

Evidence: full suite 461 passed / 5 skipped; test_asr_incremental 31/31 incl. 10 PlanSegmentsTests; ruff check clean; compileall clean. (Note: ruff format --check flags a pre-existing line in _worker_loop unrelated to this diff.)

Two commits, each independently reviewable: the behavior-preserving refactor, then the new feature.