Add torch.cuda rng state to seed save/load

[Anner-deJong]

What does this PR do?

Add torch.cuda rng state to the default seed save/load functionality:

• [Confirmed through testing on dev GPU server] torch.get/set_rng_state() does not include torch.cuda * and torch.cuda rng is not reset
• Confirmed manually on a dev GPU server the test should do what it is supposed, actual branch + dev is on a macbook without gpu, I presume the github merge tests will cover

* Presumably cause torch.cuda is not included in the torch default_generator .set_state() and .get_state() (C implementation here)

Fixes #UNKNOWN

Does your PR introduce any breaking changes? If yes, please list them.

Not aware

Before submitting

• [No] Was this discussed/approved via a GitHub issue? (not for typos and docs)
• [Yes] Did you read the contributor guideline, Pull Request section?
• [Yes] Did you make sure your PR does only one thing, instead of bundling different changes together?
• [Yes] Did you make sure to update the documentation with your changes? (if necessary)
• [Yes] Did you write any new necessary tests? (not for typos and docs)
• [Yes] Did you verify new and existing tests pass locally with your changes?
• [UNAWARE OF ANY] Did you list all the breaking changes introduced by this pull request?
• [Yes] Did you update the CHANGELOG? (not for typos, docs, test updates, or minor internal changes/refactors)

PR review

Anyone in the community is welcome to review the PR.
Before you start reviewing, make sure you have read the review guidelines. In short, see the following bullet-list:

• [Yes] Is this pull request ready for review? (if not, please submit in draft mode)
• [Mostly] Check that all items from Before submitting are resolved
• [Yes] Make sure the title is self-explanatory and the description concisely explains the PR
• [Added by @awaelchli, thanks!] Add labels and milestones (and optionally projects) to the PR so it can be classified

Did you have fun?

YES LIGHTNING IS AWESOME ⚡

Make sure you had fun coding 🙃

[Anner-deJong]

Only issue I can come up with if cuda is not available, torch.cuda returns None?
Found the tests (I forked too small part of the repo, will check + maybe update)
Putting PR in draft mode for the time being

[codecov]

Codecov Report

Merging #14384 (943e1fd) into master (34f9883) will increase coverage by 15%.
The diff coverage is 100%.

❗ Current head 943e1fd differs from pull request most recent head 1ed4258. Consider uploading reports for the commit 1ed4258 to get more accurate results

@@            Coverage Diff            @@
##           master   #14384     +/-   ##
=========================================
+ Coverage      61%      76%    +15%     
=========================================
  Files         332      332             
  Lines       26848    26892     +44     
=========================================
+ Hits        16419    20430   +4011     
+ Misses      10429     6462   -3967     

[Anner-deJong]

Only issue I can come up with if cuda is not available, torch.cuda returns None? Found the tests (I forked too small part of the repo, will check + maybe update) Putting PR in draft mode for the time being

Updated the test and did some interactive testing, ready for review. If torch is not compiled with cuda, torch.cuda.get/set.. doesnt run into any error. On the test side, torch.cuda.FloatTensor(2).normal_(), which is not guarded by an torch.cuda.is_available()

[awaelchli]

Great!

[awaelchli]

Could you also add a changelog entry? <3

[Anner-deJong]

Could you also add a changelog entry? <3

Done!