Feat/minimax m2.5 support

[xs1997zju]

Summary

Add full integration for MiniMax-M2.5 (256 experts, top-8 routing), including:

• Model spec plugin (slime_plugins/models/minimax_m2.py): Custom SelfAttention with full-dimension QK Norm (RMSNorm over all heads concatenated, with TP gather/scatter)
• mbridge weight bridge (slime_plugins/mbridge/minimax_m2.py): HF ↔ Megatron weight mapping extending Qwen2MoEBridge
• Megatron-to-HF converter (slime/backends/megatron_utils/megatron_to_hf/minimax_m2.py): Reverse conversion for saving trained checkpoints back to HF format
• Shell scripts (scripts/): Model architecture args, RL training launch script, and 3-script HF ↔ Megatron weight conversion pipeline

Key architecture differences from standard Qwen2MoE

Feature	MiniMax-M2.5	Qwen2MoE
MoE prefix	block_sparse_moe (w1/w2/w3)	mlp
QK Norm	Full-dimension (all heads concat)	Per-head
Router	Sigmoid + e_score_correction_bias	Softmax
RoPE	Partial (50%)	Full
Experts	256 × 62 layers	varies

[zhuzilin]

Thank you for the PR! Just a quick check — could you provide some evidence that the implementation is working correctly, such as W&B screenshots, logs, or validation outputs?

[xs1997zju]

Thank you for the PR! Just a quick check — could you provide some evidence that the implementation is working correctly, such as W&B screenshots, logs, or validation outputs?
@zhuzilin

aime eval results with dapo-math-17k train parameters:

`ROLLOUT_ARGS=(
--prompt-data dapo-math-17k.jsonl
--input-key prompt
--label-key label
--apply-chat-template
--rollout-shuffle
--rm-type deepscaler
--num-rollout 3000
--rollout-batch-size 32
--n-samples-per-prompt 8
--rollout-max-response-len 16384
--rollout-temperature 1
--global-batch-size 256
)

EVAL_ARGS=(
--eval-interval 20
--eval-prompt-data aime-2024.jsonl
--n-samples-per-eval-prompt 4
--eval-max-response-len 16384
--eval-top-p 1
)`