refactor: standardize packed 3d inference

- keep training attention on dense 4d tensors
- use packed 3d tensors with KV cache for inference
- extend CUDA rotary embedding to packed 3d inputs
- adapt torch, CUDA and FlashAttention backend dispatch

Benchmark: NVIDIA L20, BF16, 1B model, paged KV cache, CUDA Graph, prompt 512, generation 128 (median of 3 alternating runs)
- batch 1: 234.5 -> 242.6 tok/s (1.034x, +3.4%)
- batch 8: 1243.1 -> 1286.6 tok/s (1.035x, +3.5%)

csrc/kernels/attn_common.h

@@ -36,6 +36,8 @@ struct AttentionParams {
     const T* __restrict__ q_ptr;
     const T* __restrict__ k_ptr;
     const T* __restrict__ v_ptr;
+    const T* __restrict__ new_k_ptr;
+    const T* __restrict__ new_v_ptr;
     T* __restrict__ o_ptr;
     const bool* __restrict__ mask;
 
@@ -50,6 +52,9 @@ struct AttentionParams {
     int kv_l_stride;
     int kv_d_stride;
 
+    int new_kv_b_stride;
+    int new_kv_h_stride;
+
     int mask_b_stride;
     int mask_h_stride;
     int mask_l_stride;