[model] Add Megatron-Bridge support for MiniMax M3

[cuichenx]

Summary

Track Megatron-Bridge support for MiniMax M3 — the successor to MiniMax-M1 / MiniMax-M2 (#1475, closed).

Scope

• Model provider: configure Megatron Core with MiniMax M3 architecture (likely hybrid attention / lightning attention pattern carried over from M1/M2)
• Bridge implementation: bidirectional HuggingFace ↔ Megatron weight mapping (@register_bridge)
• Recipe: pretraining / SFT recipe under src/megatron/bridge/recipes/minimax/
• AutoBridge registration matrix coverage
• Unit tests under tests/unit_tests/models/minimax/
• Functional / convergence test with a public M3 checkpoint
• Docs page under docs/ following the model-family taxonomy

Context

• Predecessor tracking issue: Minimax M2 Model Support #1475 (MiniMax M2, closed)
• Related cross-model work that may affect this:
  • Migrate to MCore HybridModel — PoR: migrate to MCore HybridModel #4510 (M-series models are hybrid attention candidates)
  • Model support verification levels — POR: Track model support verification levels #4509
• MiniMax MoE is also listed as a NVFP4 QAD target in PoR: [support] NVFP4 Qauntization-Aware Distillation (QAD) for the following models #4512 — alignment may be needed once M3 lands.

Open questions

• HuggingFace repo / config availability for M3 (waiting on public release)
• Architectural deltas vs M2 (attention pattern, MoE topology, RoPE variant)
• Target hardware (H100 / GB200 / B200/300) and parallelism plan

References

• Bridge release tracker: MB-585 (26.08 PoR)