fix: MiniMax-M3 streaming parser when tool calls start before </mm:think>

[jkyamog]

MiniMax-M3 can start a tool call while its <mm:think> block is still
unterminated in the streamed text. The existing parser treated the open thought
as normal reasoning only while waiting for </mm:think>, which allowed the
MiniMax tool-call namespace marker to be consumed as reasoning/content in some
streaming cases.

This PR updates the dedicated MiniMax-M3 parser so that, when tools are enabled,
the reasoning parser stops at either:

• </mm:think>
• ]<]minimax[>[<tool_call>

That keeps partial reasoning quarantined during streaming and prevents MiniMax
tool-call markers from leaking into reasoning_content or visible content.

[jkyamog]

convert to draft for now, saw another case to improve the parser. doing some more testing and will push a better version.

[jkyamog]

Ok some improvements and more testing, so far the nested thinking and tools are all working well. Ready for review.

[jkyamog]

@firecoperana we caught another issue with m3 this the 2nd commit (ffe7034)

Codex explanation, I did ask if this low risk and code path is m3 specific.

This handles another MiniMax-M3 streaming edge case seen in real usage. MiniMax can begin emitting a new namespaced tool marker before the previous string tool argument is cleanly closed; without
this guard, that marker can be parsed as literal bash command text.

Risk should be low because the new logic is only called from common_chat_peg_minimax_m3_mapper::from_ast(), so it is specific to the MiniMax-M3 parser path. Other chat formats do not use this
mapper. I also added a regression test based on the observed failure shape, and test-chat-auto-parser passes.