Error only if autocast actually enabled (#96097)
I am trying to use bfloat16 AMP on a range of devices, using the `enabled` argument to actually enable/disable AMP, like this:
```python
with torch.cuda.amp.autocast(enabled=use_amp, dtype=torch.bfloat16):
```
However, this raises a RuntimeError even if enabled=False.
```
File "/venv/lib/python3.8/site-packages/torch/amp/autocast_mode.py", line 221, in __init__
raise RuntimeError('Current CUDA Device does not support bfloat16. Please switch dtype to float16.')
RuntimeError: Current CUDA Device does not support bfloat16. Please switch dtype to float16.
```
Pull Request resolved: https://github.com/pytorch/pytorch/pull/96097
Approved by: https://github.com/ngimel, https://github.com/kit1980
