[Model] Add MiniMax M3 support

[youkaichao]

Summary

• Add MiniMax M3 model support across config, processors, model registry, AMD/NVIDIA model implementations, MTP, sparse attention, and warmup paths.
• Add MiniMax M3 reasoning and tool parsers, including Rust frontend registrations and Python-facing parser wrappers.
• Add supporting kernels, quantization paths, router GEMM shape support, and targeted tests.

Duplicate-work check

• Open PR searches for MiniMax M3 and minimax_m3 found no duplicates. Broader M3 model results were unrelated.

FIX #45360

Tests

• cargo fmt --manifest-path rust/Cargo.toml --all -- --check
• cargo test --manifest-path rust/Cargo.toml -p vllm-reasoning-parser -p vllm-tool-parser -p vllm-chat

Notes

• AI assistance was used to prepare this one-commit release branch and resolve conflicts against current main.

[mergify]

Documentation preview: https://vllm--45381.org.readthedocs.build/en/45381/

[gaby]

@youkaichao The model is not working without these two PRs:

• [BugFix] Fix MXFP8 checkpoint loading when quant_method collides with online shorthand #45582
• [Bugfix] Parse MiniMax M3 streaming reasoning by text markers #45718

[fxmarty-amd]

In case a performant mixed-precision BF16-MXFP8 GEMM/MOE is not available and memory overhead (~doubling memory occupied by model weights here), it can make sense to dequantize ahead of time, hence why it was proposed previously in #35855 for MXFP4/MXFP6 (but rejected back then).

cc @mgoin fyi

I think the code fragmentation between MXFP4/MXFP6/MXFP8 emulation logic could be reduced in the future

e.g. https://github.com/vllm-project/vllm/blob/main/vllm/model_executor/layers/fused_moe/experts/mxfp8_emulation_moe.py and https://github.com/vllm-project/vllm/blob/main/vllm/model_executor/layers/fused_moe/experts/ocp_mx_emulation_moe.py