Expose `cudaMemGetInfo` (#58635)
Summary:
This PR resolves the second issue outlined in https://github.com/pytorch/pytorch/issues/58376, which has previously been discussed in https://github.com/pytorch/pytorch/issues/50722.
`cudaMemGetInfo` is bound/exposed to the Python API. An example function call is provided below:
```
device_free, device_total = torch.cuda.mem_get_info(torch.device('cuda:0'))
print(device_free, device_total)
```
In `CUDACachingAllocator.cpp`, in constant to my initial PR, the newly defined function `std::pair<size_t, size_t> raw_cuda_mem_get_info(int device)` has been moved from the `CUDACaching` namespace to the `cuda` namespace. In addition, as suugested by ezyang, `det` has been removed from all function names.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/58635
Reviewed By: zou3519
Differential Revision: D28649093
Pulled By: ezyang
fbshipit-source-id: d8b7c53e52cf73f35495d8651863c5bb408d7a6a
