Add repetition detector for degenerate generation loops

[Thump604]

Summary

Detect and stop degenerate repeating token patterns during generation. On large models (122B at 40 tok/s), a stuck loop wastes up to 13 minutes before hitting max_tokens. Addresses #65.

How it works

RepetitionDetector uses a sliding window (200 tokens) and checks every 20 tokens for patterns of length 2-50 repeated 3+ times consecutively. When detected, generation stops with finish_reason="stop" and a warning is logged.

Enabled via --repetition-detector CLI flag. Disabled by default — zero overhead when not enabled.

Changes

• New: vllm_mlx/repetition_detector.py — lightweight detector class (~70 lines)
• engine/simple.py: repetition_detector param + integration in stream_generate() loop
• server.py: Thread parameter through load_model()
• cli.py: Add --repetition-detector flag

Design decisions

• Opt-in: No behavior change without the flag
• Lightweight: O(window) check every 20 tokens, not every token
• Token-level: Uses token IDs when available, falls back to text hash
• Graceful: Sets finish_reason="stop", doesn't crash or truncate

Test plan

• Verify detection triggers on synthetic repeating patterns
• Verify no false positives on normal generation
• Verify zero overhead when disabled
• Verify MLLM path unaffected (detector is in stream_generate only)

[Thump604]

Rebased as fresh branch against current main. Reopening as new PR.