Migrate from gpus flag (to be deprecated in PL 1.7) to accelerator and devices

[nathanpainchaud]

Link to the PR that deprecated these flags, which got released in Lightning 1.7.0.

@lemairecarl and @ThierryJudge I'm tagging you as reviewer not because I expect a lot of feedback, but more as a warning about the change, since it's one of the few Trainer flags that I think you might use on your end.

[ThierryJudge]

What will be the default behaviour now ?

[nathanpainchaud]

With this config the default behavior should remain the same. accelerator='auto' will chose 'gpu' if one is available (or other types of accelerators like TPUs, etc., but will fallback to CPU otherwise. And devices is not set, by default it will try to use every available acccelerator on the machine. I did wonder if we would want to set devices to 1 to only use a single GPU in case of multi-GPU setup (like Beluga). But this is also something we could override in experiments configs that target clusters (since I'm already overriding some otherwise generic parameters like num_workers, it wouldn't be unheard of).

[nathanpainchaud]

@ThierryJudge, since you've not continued the discussion, I'll assume you've agreed to the slight change in default behavior, and I'll merge this PR later today unless I hear back from you in the meantime.