Show friendly error message when forgetting `init` in `torch.cuda` (#72404)
Summary:
# Problem
The error message `RuntimeError: Invalid device argument` is not friendly when users just forget calling `torch.cuda.init()`.
This error message is shown for example by calling `torch.cuda.reset_accumulated_memory_stats`, or other methods which internally calls [assertValidDevice](https://github.com/pytorch/pytorch/blob/6297aa114f8278f5cf5e661cba9bcf5c83f2ac1e/c10/cuda/CUDACachingAllocator.cpp#L1561-L1566).
# Reproduce
```python
$ python
Python 3.8.6 (default, Apr 1 2021, 08:23:31)
[GCC 7.5.0] on linux
Type "help", "copyright", "credits" or "license" for more information.
>>> import torch.cuda
>>> torch.cuda.reset_accumulated_memory_stats(0)
Traceback (most recent call last):
File "<stdin>", line 1, in <module>
File "/usr/local/lib/python3.8/site-packages/torch/cuda/memory.py", line 219, in reset_accumulated_memory_stats
return torch._C._cuda_resetAccumulatedMemoryStats(device)
RuntimeError: Invalid device argument.
>>> torch.cuda.current_device()
0
```
# This PR
Shows better error message like `RuntimeError: Invalid device argument 0: did you call init?`. I cited the error message from https://github.com/pytorch/pytorch/blob/6297aa114f8278f5cf5e661cba9bcf5c83f2ac1e/c10/cuda/CUDACachingAllocator.cpp#L1392-L1396.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/72404
Reviewed By: mruberry
Differential Revision: D34063268
Pulled By: ngimel
fbshipit-source-id: 0775d9c83a4a0eb0eb41bf6efecca94a00692141
(cherry picked from commit 07a1a3d0b41e1898d0c293ca77ca712de91df51e)
