MiniMax-M3 Native Tool-Call Parser

[jkyamog]

This adds native MiniMax-M3 chat-template parsing support for tool calls and reasoning extraction.

MiniMax-M3 uses its own namespace token and XML-like function-call format:

]<]minimax[>[<tool_call>
]<]minimax[>[<invoke name="tool-name">
]<]minimax[>[<arg>value]<]minimax[>[</arg>
]<]minimax[>[</invoke>
]<]minimax[>[</tool_call>

The previous PR #1963 used a MiniMax-M2-style <parameter name="..."> format. That was parser-friendly, but it did not match the native MiniMax-M3 template and caused fragile downstream behavior with clients expecting native M3 output.

Changes

• Add a specialized MiniMax-M3 PEG parser in common/chat.cpp.
• Detect MiniMax-M3 templates by the native namespace/tool-call markers.
• Parse <mm:think>...</mm:think> into reasoning_content.
• Parse single and parallel native tool calls into structured tool_calls.
• Remove the obsolete models/templates/MiniMax-M3.jinja override, which used the old M2-style compatibility wrapper.
• Add a regression test covering reasoning plus two native tool calls.

Validation

• Syntax-checked common/chat.cpp and tests/test-chat-template.cpp against two existing compile databases.
• Built successfully in the Docker CUDA image used by the local deployment.
• Deployed locally with MiniMax-M3 Q6 and verified an OpenAI-compatible tool-call smoke:
  • content was empty
  • reasoning_content contained hidden reasoning
  • tool_calls contained a structured function call with JSON arguments
  • used Hermes and Pi Agent as end to end test

Notes

This intentionally relies on the native MiniMax-M3 chat template, normally embedded in the GGUF metadata. No external --chat-template-file override is required when the model already carries the native template.

This is a breaking cleanup for the old PR #1963 MiniMax-M3 template override. That file preserved a MiniMax-M2-style <parameter name="..."> wrapper and should not be kept now that native M3 tool parsing is supported directly.

The parser handles MiniMax-M3 generation prompts that already end with <mm:think>, so generated reasoning is still extracted into reasoning_content instead of leaking into visible content.

[jkyamog]

the short story of this PR, Minimax M3 was very unreliable with tool calls. Initially I thought it was the model, that is why I switched to GLM-5.2. However I thought surely Minimax M3 Q6 should not be worse the smaller Qwen 27B, 35B. After asking and pushing codex, eventually I understood that the PR $1963 last week was problematic. Sorry breaking change, codex insisted on keeping the template but I think that is confusing having a template that is not really used most of the time.