refactor: unify paged and contiguous attention kernels via KVSource policy

- merge AttentionParams and PagedAttentionParams into one struct
- add attn_kv_source.cuh with ContigKV/PagedKV addressing policies
- template prefill/decode kernels (MMA + scalar) on the KV policy, deleting the four duplicated attn_paged_*.cuh variants
- template dispatcher launchers on KV; single combine kernel
- verify: all correctness tests pass and SASS matches baseline (no perf regression)

csrc/tests/test_utils.cuh

@@ -120,7 +120,7 @@ inline void set_default_strides(P& p) {
     p.mask_q_stride = 0;
 }
 
-// Set default Q strides for contiguous b h l d layout on PagedAttentionParams.
+// Set default Q strides for a paged decode params struct.
 template<typename P>
 inline void set_default_paged_strides(P& p) {
     p.q_stride_b  = p.q_head * p.q_len * p.head_dim;