Add MiniMax-M3 support for GRPO training

## Description

Add support for MiniMax-M3 in NeMo RL, building on the MiniMax-M2.7 support from #2251 and #2685.

MiniMax-M3 introduces several architectural differences from M2.7, including:

- Native multimodal input support
- MiniMax Sparse Attention (MSA)
- Approximately 428B total parameters with approximately 23B activated parameters
- Support for up to 1M-token context
- Reasoning modes: enabled, adaptive, and disabled

The initial goal is to enable reproducible GRPO training and evaluation for the
`MiniMaxAI/MiniMax-M3` checkpoint.

## Scope

- Add MiniMax-M3 model loading and training support.
- Support distributed training with the appropriate AutoModel backend.
- Support MoE parallelism and the M3 attention implementation.
- Add rollout/generation support, including the M3 tokenizer, chat template,
reasoning behavior, stop-token handling, and response-mask construction.
- Add a runnable GRPO configuration for MiniMax-M3.
- Validate checkpoint save/reload and generation parity.
- Document supported hardware, parallelism configuration, precision, and
maximum validated sequence length.
- Preserve existing MiniMax-M2.7 support.

## Initial acceptance criteria

- [ ] `MiniMaxAI/MiniMax-M3` can be loaded successfully by NeMo RL.
- [ ] A minimal GRPO run completes end-to-end on the target NVIDIA cluster.
- [ ] Training loss, reward, response length, and KL metrics are reported.
- [ ] Checkpoint save/reload succeeds and produces numerically consistent output.
- [ ] A documented recipe is added under `examples/configs/`.
- [ ] A short validation run demonstrates non-zero learning signal.
- [ ] Configuration and documentation identify any limitations around:
multimodal inputs, MSA, long-context training, precision, or rollout backend.
- [ ] Existing MiniMax-M2.7 tests and recipes continue to pass.

## Validation

Please include:

- Tested model/checkpoint revision
- GPU type, number of nodes, and parallelism settings
- Precision and checkpoint format
- Validated sequence lengths
- Rollout backend and version
- Initial GRPO reward/accuracy curves
- Throughput and memory measurements
- Known limitations or unsupported features

Related work:

- #2251 — MiniMax-M2.7 support
- #2685 — MiniMax-M2.7 implementation
- MiniMax-M3 model: https://huggingface.co/MiniMaxAI/MiniMax-M3
- MiniMax-M3 technical report: https://arxiv.org/abs/2606.13392