settings: Add ORT thread count setting with auto-tune benchmark.

[andrewleech]

Summary

ORT's default of "use all cores" isn't always the fastest for ONNX inference — on my Zen 5 machine, 6 threads gives 7.1x realtime vs ~6.7x at 16 threads for Qwen3 0.6B Int4. But finding the optimal count requires manual trial-and-error with no feedback.

This adds two things:

1. Thread count setting — a number input (0-32, where 0 = auto) in Advanced → Experimental Features that controls transcribe_rs::accel::set_ort_intra_threads. Changing it unloads the current model so the next transcription picks up the new session configuration.

2. Auto-tune benchmark — an "Auto" button next to the input that opens a modal dialog, transcribes real recordings from the user's history at different thread counts, and shows the results. It tests a coarse grid (1, 2, 4, 6, 8, 12, 16, 20, 24, 32), starting from mid-range for quick initial feedback. On completion it shows a table of all results with RTF (real-time factor) so the user can see the performance curve and apply the best value.

The benchmark:

• Selects recordings from history totalling ≥15 seconds (longest-first)
• Auto-loads the selected model if it was idle-unloaded
• Tests mid-range first, then max, then 1, then fills the remaining grid
• Supports cancellation between trials
• Uses RAII to guarantee the original thread count is restored on completion or panic
• Shows all trial results in a table with the best highlighted

Depends on transcribe-rs feat/ort-thread-count (PR forthcoming). The DROP commit pins to this branch and should be dropped when it merges upstream.

Testing

• Tested on Windows (Zen 5 + AMD 860M iGPU) with Qwen3 ASR 0.6B Int4 and Parakeet V3
• Verified: setting persists across restarts, model reloads on change, benchmark completes and applies result
• Verified: cancel mid-benchmark restores previous state
• Verified: benchmark with no recordings shows error message

Trade-offs and Alternatives

• The benchmark requires a full model reload per trial (~4-12s each depending on model size) because ORT session thread count is baked in at creation time. For a 10-point grid this means ~1-2 minutes total.
• The coarse grid [1,2,4,6,8,12,16,20,24,32] may miss the exact optimum but gets within ±2 threads of it.
• The benchmark locks the engine exclusively — the user can't record while it runs.

[andrewleech]

Depends on transcribe-rs PR #64 for set_ort_intra_threads / get_ort_intra_threads. The DROP commit pins to that branch and should be dropped when it merges.