feat: add vLLM 0.16 mistral option parity for exllamav3

[lesj0610]

Summary

• implement tool_call_parser=mistral for both Mistral formats
  • legacy JSON list: [TOOL_CALLS] [{"name":...,"arguments":...}]
  • v11+ segmented format: [TOOL_CALLS]name{...}[TOOL_CALLS]...
• wire parser dispatch and native-generation behavior for mistral parser key
• add tokenizer_mode model config/plumbing and expose it in model metadata
• apply mistral-safe tool call id normalization/truncation in chat formatting path
• improve mistral reasoning parser token handling with mistral_common special tokens fallback
• keep non-mistral models on existing/default paths (fallback-safe behavior)

Files / Areas

• endpoints/OAI/utils/tools.py
• endpoints/OAI/utils/chat_completion.py
• endpoints/OAI/utils/parser_options.py
• common/config_models.py
• endpoints/core/types/model.py
• backends/exllamav3/model.py
• backends/exllamav2/model.py
• docs + tests updates

Validation

• PYTHONPATH=. pytest -q tests/tool_parser_test.py tests/parser_options_test.py tests/mistral_reasoning_parser_test.py tests/exaone4_reasoning_parser_test.py tests/wheel_test.py
• result: 34 passed
• note: full pytest -q in this environment still fails in tests/model_test.py due missing exllamav2 package (ModuleNotFoundError), unrelated to this change set.

Related

• includes prior prerequisite merges/cherry-picks discussed for OpenAI compatibility: reasoning parser coverage and robust tool-call handling #411 and Full tool-calling support, inference abort fixes, XML parsing, OpenAI streaming compliance #413 alignment before this feature work.