fix: harden streaming reasoning/tool parsing across engines

[krystophny]

This builds on waybarrios#177 and adds the remaining hardening that showed up under broader local + live validation.

What this adds on top of the branch:

• preserve raw tool-call text for non-streaming chat in both SimpleEngine and BatchedEngine
• parse Qwen function/parameter XML in addition to the JSON and bracket formats
• make Qwen/Harmony streaming completion checks use accumulated text instead of depending on a lucky final delta
• route GPT-OSS streaming through the split-token Harmony state machine
• dedupe repeated identical Harmony commentary calls and emit only newly completed calls in streaming
• treat atomic tool-call parsers as terminal in streaming, so completed tool calls finish with finish_reason="tool_calls"
• replace the format-specific end-of-stream tool fallback with a validated final parse over accumulated tool text
• add direct streaming and engine-preservation coverage for these cases

Validation on current main before splitting this out:

• pytest -q tests -x --timeout=120 -m 'not slow'
• result: 1090 passed, 11 skipped, 18 deselected
• live E2E passes:
  • Qwen3.5-2B simple + batched
  • Qwen3.5-9B simple + batched
  • Qwen3-Coder-30B simple
  • GPT-OSS-20B with gpt_oss reasoning parser, simple + batched
  • GPT-OSS-20B with harmony reasoning parser, simple
  • Devstral-Small-2 simple + batched with forced named tool choice

Validation on this branch:

• python -m pytest -q tests/test_harmony_parsers.py tests/test_tool_parsers.py tests/test_simple_engine.py tests/test_engine_tool_output_preservation.py
• python -m py_compile vllm_mlx/server.py vllm_mlx/engine/simple.py vllm_mlx/engine/batched.py vllm_mlx/reasoning/gpt_oss_parser.py vllm_mlx/reasoning/harmony_parser.py vllm_mlx/tool_parsers/harmony_tool_parser.py vllm_mlx/tool_parsers/qwen_tool_parser.py

Note: importing server.py on the base branch still hits pre-existing Responses-API type references that are unrelated to this streaming/tool fix, so I kept the branch-local rerun focused on the directly touched parser/engine coverage.