Fix LFM2.5 pythonic tool parser auto-detection

[ChristianWeyer]

What

_infer_tool_parser() in tokenizer_utils.py detects the pythonic tool-call
format via the chat-template token <|tool_list_start|>. LFM2.5 renamed those
delimiters to <|tool_call_start|> / <|tool_call_end|> (the inner pythonic
call format is unchanged), so the detection misses for every LFM2.5
checkpoint — including the official LiquidAI/LFM2.5-8B-A1B,
LiquidAI/LFM2.5-8B-A1B-MLX-8bit, LiquidAI/LFM2.5-1.2B-Thinking, etc.

Symptom: loading any LFM2.5 model gives tokenizer.has_tool_calling == False,
so mlx_lm.server's /v1/chat/completions returns tool calls as raw
<|tool_call_start|>[func(arg='val')]<|tool_call_end|> text inside
message.content, with tool_calls = [] and finish_reason = "stop".
OpenAI-compatible clients (Mastra, LangChain, ai-sdk, …) see no tool call and
the agent loop breaks.

Fix

Match either token. The pythonic parser already handles both formats — only
the auto-detection needed updating.

elif (
    "<|tool_list_start|>" in chat_template
    or "<|tool_call_start|>" in chat_template
):
    return "pythonic"

Same root cause and fix shape as the parallel llama.cpp issue
ggml-org/llama.cpp#23838,
and the same class of token-rename detection lag that prompted the Gemma 4
auto-detect addition.

Test

Added TestTokenizers.test_infer_tool_parser in tests/test_tokenizers.py
covering:

• LFM2 original template (<|tool_list_start|>) — still returns "pythonic"
• LFM2.5 template (<|tool_call_start|>) — now returns "pythonic" (was None)
• Negative path (no tool tokens, and None input)

This is a pure unit test against _infer_tool_parser, so no HF download —
fast and deterministic. Verified locally that the new test fails on main
without the fix and passes with it.

End-to-end verified against the real LiquidAI/LFM2.5-8B-A1B-MLX-8bit
tokenizer:

>>> from mlx_lm.utils import load_tokenizer
>>> tok = load_tokenizer("LiquidAI/LFM2.5-8B-A1B-MLX-8bit")
>>> tok.has_tool_calling
True
>>> tok.tool_call_start, tok.tool_call_end
('<|tool_call_start|>', '<|tool_call_end|>')

And via mlx_lm.server /v1/chat/completions with a tools payload — after
the fix, the response now correctly returns:

{
  "finish_reason": "tool_calls",
  "message": {
    "tool_calls": [{
      "function": {"name": "...", "arguments": "{\"query\": \"...\"}"},
      "type": "function",
      "id": "..."
    }]
  }
}

All existing tests/test_tool_parsing.py and tests/test_tokenizers.py tests
still pass; ran black (25.1.0) and isort --profile=black per
CONTRIBUTING.md.