[Re-land] [CUDA graphs] Clear autocast amp cache (#81896)
Re-lands #81558 that got reverted due failing tests.
This failure happened because of the test that I poorly designed. [The loop here](https://github.com/pytorch/pytorch/pull/81558/files#diff-893b1eea27352f336f4cd832919e48d721e4e90186e63400b8596db6b82e7450R3837) is doing `cache_enabled=False` and then `cache_enabled=True`. By doing this loop the graph from previous iteration (case `False`) conflicts with the next one (case `True`). I redesigned the test such that it does not do any loops. The new test does separate function calls with different argument values.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/81896
Approved by: https://github.com/ngimel
