[CUDA] Recycle CUDA events (#2604)

* Make CudaEvent a CudaHandle

* Add caching for CudaEvent

* Make sure cuda events are destroyed at last

* Fix headers

* SharedEvent => AtomicEvent

* RawCudaEvent => CudaEventHandle, CudaEventWrapper => CopyableCudaEvent

* Remove unneeded asserts

mlx/backend/cuda/eval.cpp

@@ -15,8 +15,9 @@ bool is_available() {
 }
 
 void new_stream(Stream s) {
-  // Force initalization of cuda, so cuda runtime get destroyed at last.
-  cudaFree(nullptr);
+  // Force initalization of CUDA by creating an event, so the CUDA runtime and
+  // our CUDA event pool get destroyed last.
+  cu::CudaEvent(cudaEventDefault);
   // Ensure the static stream objects get created.
   cu::get_command_encoder(s);
 }