Add rope_norm_store_kv fusion op

[xueyangcs]

This PR adds a fused CUDA operator that performs RoPE rotation, optional QK RMSNorm, and blocked KV-cache write in a single kernel.

API

API	Input dtype	Output	Description
hpc.rope_norm_store_kv	BF16	out_q; K/V written in-place to KV cache	Standard BF16 inference
hpc.rope_norm_store_kv_fp8	BF16 → FP8	(out_q_fp8, q_scale, split_k_flag); K/V written to FP8 KV cache	FP8 quantized inference

Both variants support prefill and decode modes via is_prefill.

Policy Parameters

qk_norm_policy

Controls whether RMSNorm is applied to Q/K and its order relative to RoPE.

Value	Behavior
0	No RMSNorm
1	RoPE → RMSNorm
2	RMSNorm → RoPE

quant_policy

Controls the Q quantization granularity. K/V always use static scaling via k_scale / v_scale.

Value	Name	Q Quantization
1	dqskv	Dynamic per-token per-head; scale computed by the kernel and written to q_scale
2	sqskv	Static; uses the caller-supplied q_scale_inv