[WIP] feat(dcu): add BW1000 DCU backend support

[chenyou74]

Summary

This PR adds Hygon BW1000/DCU backend support.
Main changes:

• Add USE_DCU / USE_RCCL build options.
• Add DCU runtime and device guard implementation.
• Add RCCL communication backend.
• Add DCU HIP kernels.
• Enable GPT-2 and LLaMA3 examples to run with --device dcu.

FM9G text-only LoRA support

This PR additionally adds:

• FM9G text-only model configuration and attention support
• Hugging Face FM9G checkpoint conversion and loading
• LoRA training for q_b_proj, kv_b_proj, and o_proj
• FP32 and BF16 training support on BW1000/DCU

Scope

The current FM9G integration covers the text-only LLM path and LoRA
training. The vision encoder, resampler integration, image preprocessing,
and multimodal training are not included.