Let ck_splitk_decoder to use ck_tile headers only

[qianfengz]

This PR solves a xformers building issue by removing the ck_splitkv_decoder's depending on the old_ck headers, making ck_splitkv_decoder only uses the ck_tile headers.

Without the fixing from this PR, building will fail while updating the ck_tiled submodule to latest ck develop due to the conflicts from the old_ck and ck_tile header files.

This is a quick fix, so performance is not considered. To achieve to the same performance, the inner_product in ck_tile_attention_inner_product should be improved.