fix(vllm): non-streaming tool-call regression after #10351

[pos-ei-don]

Summary

Non-streaming chat.completions with tools=[...] silently drops the tool
call (finish_reason: "stop", empty content, no tool_calls[]) when a
tool_parser is configured. Streaming requests are unaffected.

Root cause

#10351 introduced native streaming via parser.extract_tool_calls_streaming
and gated the post-loop extract_tool_calls block on
native_streaming and not native_streaming_error.

native_streaming is a capability flag — it becomes True as soon as
tp_request builds and the parser exposes extract_tool_calls_streaming.
It is not a state flag telling us the per-delta loop actually ran.

For non-streaming requests it is still True, so:

• the if branch (extract_tool_calls on the full text) was skipped
• the elif branch (content = "", "the stream already emitted everything")
  fired

The result: the tool call is neither extracted nor streamed — it's lost.

Fix

Add an explicit streaming and … guard on both branches so
extract_tool_calls runs for the non-streaming path (and for streaming
paths that fell back to the buffered mode). No change to streaming
requests where native streaming ran cleanly.

Reproduction

Model: saricles/Qwen3-Coder-Next-NVFP4-GB10, vLLM 0.24.0, LocalAI
v4.5.6, coder.yaml with options: - tool_parser:qwen3_coder.

curl -s -X POST http://localhost:8080/v1/chat/completions \
  -H 'Content-Type: application/json' \
  -d '{"model":"coder","stream":false,
       "messages":[{"role":"user","content":"7*8 with the calc tool"}],
       "tools":[{"type":"function","function":{"name":"calc",
         "parameters":{"type":"object",
           "properties":{"expression":{"type":"string"}}}}}]}'

Before this patch: finish_reason: "stop", content: "", tool_calls: [].

After this patch: finish_reason: "tool_calls",
tool_calls[0].function.name: "calc",
arguments: '{"expression":"7*8"}'.

Streaming (stream: true) re-verified after the change: delta.tool_calls
arrives token-by-token, finish_reason: "tool_calls", no raw XML in
content (behaviour from #10351 unchanged).

Scope

Two lines of behaviour change plus a comment explaining the capability-vs-state
distinction so the guard isn't accidentally removed again.