Cast to >=float32 tensor when passing scalar to self.log

[MF-FOOM]

What does this PR do?

Ensures that values get cast to float32+ tensors when passed to self.log.

Fixes #18984

Before submitting

• Was this discussed/agreed via a GitHub issue? (not for typos and docs)
• Did you read the contributor guideline, Pull Request section?
• Did you make sure your PR does only one thing, instead of bundling different changes together?
• Did you make sure to update the documentation with your changes? (if necessary)
• Did you write any new necessary tests? (not for typos and docs)
• Did you verify new and existing tests pass locally with your changes?
• Did you list all the breaking changes introduced by this pull request?
• 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:

Reviewer checklist

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

---

📚 Documentation preview 📚: https://pytorch-lightning--19046.org.readthedocs.build/en/19046/

[codecov]

Codecov Report

Merging #19046 (3ee07db) into master (1c86011) will decrease coverage by 27%.
Report is 55 commits behind head on master.
The diff coverage is 100%.

Additional details and impacted files

@@            Coverage Diff            @@
##           master   #19046     +/-   ##
=========================================
- Coverage      76%      49%    -27%     
=========================================
  Files         447      435     -12     
  Lines       36375    36419     +44     
=========================================
- Hits        27520    17816   -9704     
- Misses       8855    18603   +9748     

[carmocca]

I tried writing a test for this but it's complicated since the metric will be moved to float32 after, so we would need to dynamically patch the call to result.log. It's not worth it.

def test_log_tensor_wrap_dtype():
    class MyModel(LightningModule):
        def test_step(self, *_):
            self.log("foo", 1.23, batch_size=1)

    model = MyModel()
    trainer = Trainer(logger=False, enable_checkpointing=False, enable_model_summary=False, enable_progress_bar=False)
    # using the context manager class for cleanup, but users would use `set_default_dtype` directly
    with _DtypeContextManager(torch.float16):
        trainer.test(model, [0])
    assert trainer.callback_metrics["foo"].dtype == torch.float32

This attempt passes anyways without your PR checked out

LGTM