refactor: reorganize CUDA kernels into per-family directories

- move attention kernels to csrc/kernels/attention/ and rotary to rotary/
- add shared common/mma.cuh (mma_sync, ldmatrix) and device.cuh (sm checks)
- split fp8_mm into three-layer fp8/common.h, gemm.cuh, mm.cu
- fix fused FP8 GEMM ldmatrix lane indexing to fix OOB shared reads
- update extension ops, loader, and kernel tests

csrc/tests/attn_test.cu

@@ -7,7 +7,7 @@ nvcc -I csrc -arch=sm_89 -O3 \
 */
 
 #include "test_utils.cuh"
-#include "../kernels/attn_dispatchers.cuh"
+#include "../kernels/attention/dispatchers.cuh"
 
 struct DecodeDispatch { AttentionParams<bf16>& p; template<int H> void operator()() { dispatch_decode<H>(p, 0); } };
 struct PrefillDispatch { AttentionParams<bf16>& p; template<int H> void operator()() { dispatch_prefill<H>(p, 0); } };