[feat] GTP+DCP #6501

GTP+DCP(Distributed Checkpoint) has been supported in the GTP init MR #4967

This issue will keep tracking following-up bug-fixes / perf optimizations for GTP+DCP related stuffs.

• fix GTP+DCP for GDP moudle ([fix] fix GTP+DCP ckpt saving/loading for GDP module #6503 )
• fix GTP+DCP when parallel_state is not initialized (fix(gtp+dcp): elect the shard writer from the caller's process group #6775 )
• fix the case when some params is sharded across GTP while been duplicated across TP group (e.g., the latent gemm) (Route custom process groups through GDP and latent MoE checkpoints #6781 )
• GTP+DCP+fp8-param save->load is not bitwise matching
  - MXFP8 (bytes + scale_A) → dequantize → BF16 → save → load → quantize → MXFP8 (bytes + scale_B)
  - Where scale_A (from fp32 master w) ≠ scale_B
  - quantize(dequantize(mfxp8) is a lossy round-trip, not identical
  - solution: during loading: cast from fp32 main params into mxfp8 params instead (related MR: Re-derive quantized params from FP32 main params on checkpoint load #6666 )
• Fix GTP checkpoint load failing across different GTP degrees (Fix GTP checkpoint load failing across different GTP degrees #7293)