fix(server): integrate tool call parser into reasoning parser streaming path

[mxl]

Running Qwen 3.5 model results in printing tool calls as raw xml and not applying actual tools.

Summary

• run the streaming tool-call parser in the reasoning-parser path after reasoning content has been stripped from streamed output
• suppress tool-call markup from normal content chunks and emit structured tool_calls chunks when parsed tool calls are detected
• preserve the tool_calls finish reason on the final streamed chunk when generation ends with parsed tool calls

[waybarrios]

Clean fix for a real bug — tool calls were leaking as raw XML when the reasoning parser was active. The approach is correct and tests are solid.

One minor bug:

request.model vs _model_name inconsistency

The new tool_chunk in the reasoning path uses model=request.model, but every other chunk in the function (including the equivalent tool_calls chunk in the standard path) uses model=_model_name:

# New code (reasoning path):
tool_chunk = ChatCompletionChunk(
    id=response_id,
    model=request.model,      # ← client-provided value
    ...
)

# Standard path (line ~2252) and all other chunks:
chunk = ChatCompletionChunk(
    id=response_id,
    model=_model_name,         # ← actual served model name
    ...
)

These can differ when --served-model-name is set. Should be _model_name for consistency.

[waybarrios]

Pushed a small fix (1d16507) — was using request.model instead of _model_name in the reasoning tool chunk. Consistent with the rest of the function now.