fix: Improve safety of logging images with implicit formats

[natikgadzhi]

Summary

This pull request makes it so:

• If live.log_image is called without image format in the name, we will try to infer the format from the image
• If we could not infer the format, we will raise an error message early in could_log instead of falling through to PIL.Image.save

Closes #743

Changes

@dberenbaum, I felt that both approaches need to be implemented together. Inferring the format is nice, but only works in some situations and not others. So for a situation where the format could not be inferred, I added an additional check that will fail early and won't attempt saving the image.

My Python must be not optimal. The changes are split into two commits, feature-wise.

• I've also added unit tests for the behavior.
• Behavior change: in could_log I suggest we take name and val, not just val. This might be a bad idea if it's used elsewhere (not in this repository, I checked). It's not in public API, though.

I'm not strongly attached to the could_log idea. In the end of the day, it's only saving us a bit of time, and clarifies the error, but it's not strictly required.

TODO

• ❗ I have followed the Contributing to DVCLive
  guide.

• 📖 If this PR requires documentation updates, I have created a separate PR (or issue, at least) in dvc.org and linked it here.

Thank you for the contribution - we'll try to review it as soon as possible. 🙏

[natikgadzhi]

I wonder if that's worth documenting? Sort of a minor change, but happy to put a PR into the docs site, too.

[natikgadzhi]

Also, happy to write up the python docs in all the live.* functions if that's something that you think is useful.

[dberenbaum]

Thanks @natikgadzhi! Looks like it all works, but have some suggestions for refactoring it a bit.

[dberenbaum]

WDYT about handling errors here better? This would be a good place to suggest installing dvclive[image].