perf(minimax-h3): Optimize MiniMax-H3 fused QKV norm and RoPE on Intel XPU

[qiuxin2012]

Description

Optimize MiniMax-H3 QKV projection, Q/K norm, and RoPE execution with registered fused backends.

• Add the intel_xpu fused QKV norm + RoPE backend.
• Keep a generic Triton backend using TorchRealRope FP32 intermediate computation.
• Integrate RoPE into the QKV preparation path with automatic fallback.

Example:

{
  "use_fused_qkv": true,
  "use_fused_qkv_norm_rope": true,
  "qkv_norm_rope_type": "intel_xpu"
}