OpenAI compatibility: reasoning parser coverage and robust tool-call handling

[lesj0610]

Summary

This PR improves OpenAI-compat behavior in tabbyAPI (ExLlamaV3 backend path), with a focus on correctness, maintainability, and parser/tool-call robustness.

Included changes

• Added a vLLM-style reasoning parser framework and full parser registry under endpoints/OAI/reasoning/.
• Added parser selection config (reasoning_parser) with default fallback (basic) when unspecified.
• Integrated reasoning extraction into both non-streaming and streaming chat completion flows.
• Preserved strict separation between reasoning / reasoning_content and content.
• Hardened tool-call handling:
  • supports tool_choice: none | auto | required and named function choice,
  • normalizes model-emitted tool payload forms (including <tool_call>...</tool_call>),
  • aligns final response behavior with OpenAI-compatible tool_calls handling.
• Added Jinja tojson compatibility for templates using tojson(ensure_ascii=False).
• Added OpenWebUI compatibility for thinking toggles by mapping top-level enable_thinking / thinking into chat_template_kwargs (template_vars) during request validation.

Validation

• compile checks on modified modules,
• reasoning on/off and streaming checks on /v1/chat/completions,
• tool-choice path checks (none, required, named function),
• EXAONE template render path verification.

License and attribution

• Added third-party notices for vLLM-derived parser code:
  • THIRD_PARTY_LICENSES.md
  • LICENSES/Apache-2.0.txt
• Added SPDX headers to reasoning parser files where applicable.

Additional update (EXAONE4 parser hardening)

• Reworked endpoints/OAI/reasoning/exaone4_reasoning_parser.py into a standalone parser (no runtime delegation to DeepSeekR1ReasoningParser / DeepSeekV3ReasoningParser).
• Implemented EXAONE-specific behavior for enable_thinking on/off in both non-streaming and streaming flows.
• Added coverage test: tests/exaone4_reasoning_parser_test.py (prefill/no-start-token streaming, end-token split, non-thinking content-only behavior).

Additional update (vLLM-style tool parser registry)

• Added parser-key dispatch for hermes, llama/llama3_json/llama4_json, openai, pythonic, qwen3_coder, qwen3_xml, deepseek_v3, deepseek_v31, deepseek_v32
• Added native second-pass generation mode for parser families that require non-JSON syntax
• Added parser alias normalization (llama -> llama3_json) and native-generation decision helpers
• Added focused tests for parser mapping, parser-key dispatch precedence, and fallback behavior
• Head commit: b71647a