Refactors shared memory copy initialization code

Simplifies the creation of tiled shared memory copy objects by removing
the intermediate step of getting thread slices directly from the factory
function. The refactored approach creates the tiled copy object first,
then obtains the thread slice separately for better code clarity.

Author: LoserCheems

csrc/src/flash_bwd_kernel.h

@@ -370,7 +370,7 @@ inline __device__ void compute_dq_dk_dv_1colblock(const Params &params, const in
 
     // Partition sP and sdS to match the accumulator partitioning
     // This has to be tiled_mma_sdp, not tiled_mma_dkv
-    // auto smem_thr_copy_PdS = make_tiled_copy_C(typename Kernel_traits::SmemCopyAtomPdS{}, tiled_mma_sdp).get_thread_slice(tidx);
+    // auto smem_tiled_copy_PdS = make_tiled_copy_C(typename Kernel_traits::SmemCopyAtomPdS{}, tiled_mma_sdp);
     auto smem_tiled_copy_PdS = make_tiled_copy_C_warpcontiguousN<MMA_N_SdP>(typename Kernel_traits::SmemCopyAtomPdS{}, tiled_mma_sdp);
     auto smem_thr_copy_PdS = smem_tiled_copy_PdS.get_thread_slice(tidx);
     Tensor tPsP = smem_thr_copy_PdS.partition_D(sP);        // ((Atom, AtomNum), PIPE_M, PIPE_N)