[CUDA] cuDNN forward attention (#2743)

* Separate sdpa kernels in another file

* Initial support for cuDNN SDPA

* Diable a few corner cases

* Remove scaled_dot_product_attention.h

* Use cuDNN attention for prefilling

* cuDNN SDPA requires Ampere and later

* Address reviews

* Do contiguous copy of inputs

mlx/backend/cuda/CMakeLists.txt

@@ -44,6 +44,7 @@ target_sources(
           ${CMAKE_CURRENT_SOURCE_DIR}/reduce/row_reduce.cu
           ${CMAKE_CURRENT_SOURCE_DIR}/rms_norm.cu
           ${CMAKE_CURRENT_SOURCE_DIR}/rope.cu
+          ${CMAKE_CURRENT_SOURCE_DIR}/scaled_dot_product_attention.cpp
           ${CMAKE_CURRENT_SOURCE_DIR}/scaled_dot_product_attention.cu
           ${CMAKE_CURRENT_SOURCE_DIR}/scan.cu
           ${CMAKE_CURRENT_SOURCE_DIR}/slicing.cpp