Call comms / compute overlap passes when compile=False

[fmassa]

Previously, when we would call AutoParallel with compile=False, we wouldn't have any of the comms / compute overlap passes being applied to the model.

This effectively meant that we would need compile=True to have a performant autoparallelized model.

I've for now decided to call into all the post_grad passes, but it is also possible that we only call into the comms / compute overlap passes, to keep graph modifications to a minimum.

I'm now calling into the comms / compute reordering pass even when compile=False

[xmfan]

some of the post grad passes are bad for perf unless lowered e.g. view_to_reshape which materializes all views

[fmassa]

I've changed it to only call into the comms / compute reordering pass, to keep graph changes to a minimum

[wconstab]

seems OK to me. i will say that it's not super clear to me what the best formulation is. It's a little arbitrary which compiler passes to put 'inside' vs 'outside'.

from a use-case perspective, it seems nice to always have the distributed passes run, even if codegen isn't important. otoh, other things like cudagraph might also be preferred, even without codegen. For debugging, the unmodified original graphmodule might be nice to get out? (though, you can see it in its various states of transformation using tlparse).