[CUDA] Initial implementation of Convolution with cuDNN (#2385)

* Link with cuDNN

* Initial implementation

* Remove backend apis

* Fix recording cudnn conv

* More unused backend apis

* Fix C++ conv tests

* include cudnn as python dep

* Install libcudnn9-dev-cuda-12 in CI

* cudnn only accepts contiguous inputs

* Switch to backend apis

* Plan needs to be kept alive

* Turn off tf32

* Add cache

* Test the native cuda graph api

* Set cudnn stream before execution

* Make LRUCache more like a normal container

* Do error check for cublas handle

* Zero-initilizing array

* Use tf32 for conv

* Skip TestConv.test_torch_conv_2D test

---------

Co-authored-by: Awni Hannun <awni@apple.com>

python/scripts/repair_cuda.sh

@@ -16,7 +16,7 @@ rm "${repaired_wheel}"
 mlx_so="mlx/lib/libmlx.so"
 rpath=$(patchelf --print-rpath "${mlx_so}")
 base="\$ORIGIN/../../nvidia"
-rpath=$rpath:${base}/cublas/lib:${base}/cuda_nvrtc/lib
+rpath=$rpath:${base}/cublas/lib:${base}/cuda_nvrtc/lib:${base}/cudnn/lib
 patchelf --force-rpath --set-rpath "$rpath" "$mlx_so"
 python ../python/scripts/repair_record.py ${mlx_so}