Stop generation on consecutive duplicate tool calls

[chopchop-jiahao]

Summary

When using tool calling with certain models (e.g. Qwen3.5-9B), the model sometimes generates 43–64 identical tool calls in a single response until max_tokens is exhausted. This wastes compute and confuses the agent loop.

This PR adds a lightweight ToolCallDedup class that compares each completed tool call with the previous one. On consecutive duplicate, generation stops early with finish_reason=tool_calls.

Changes:

• New mlx_lm/tool_call_dedup.py — simple consecutive duplicate detector (exact string comparison, no JSON normalization)
• Integration in server.py — one instance per request, checked at each tool_call_end token
• Unit tests covering: first call, different calls, consecutive/non-consecutive duplicates, whitespace sensitivity, logging, state management
• Integration tests through the full HTTP pipeline, including a 43-duplicate reproduction of the real-world scenario, both streaming and non-streaming

Design decisions:

• Only detects consecutive duplicates (A-A), not alternating patterns (A-B-A-B) — sufficient for observed behavior and avoids false positives
• Exact text comparison — conservative, no risk of normalizing away meaningful differences
• No configuration flag — the behavior is strictly beneficial (stops degenerate loops)
• Per-request instance — no thread safety concerns

Closes #613