Remove `Event.INIT`; move `Event.TRAINING_START` to before the DDP

[ravi-mosaicml]

Once #43 is merged in, then Event.INIT can be removed.

Event.TRAINING_START can be moved up to the line right before ddp.prepare_module(). All existing algorithms that currently run on training start would be compatible with this change.

• Augmix: :heavy_check_mark:
• Colout: :heavy_check_mark:
• Curriculum Learning: @moinnadeem what is being done on training start? I do not see any inputs or outputs being preserved.
• Randaugment: :heavy_check_mark:
• Sam: :heavy_check_mark:. In fact, it might be broken as it currently runs on training_start, after optimizers have been moved onto the device
• Scale Schedule: :heavy_check_mark:
• Seq Length Warmup: @moinnadeem what is being done on training start? I do not see any inputs or outputs being preserved.
• SWA: :heavy_check_mark: This change would simplify this algorithm, as it would apply before the model is moved onto the device (and it wouldn't need to take that into account anymore)

TODO:

• [x] Implement and merge #43
• [x] Fix algorithms as described above
• [x] Refactor the training loop to support this change.
• [x] Update documentation. We frequently refer to TRAINING_START happening after the DDP fork and all DDP madness. This changes that. Will need to take a throughout read through the docs/ folder.

[jbloxham]

I'm not sure how I feel about removing INIT. I definitely agree it's hacky how Event.INIT currently needs to exist between model initialization and optimizer initialization to accommodate surgery, but #43 will allow us to fix those issues without just removing the entire event. I think INIT should be called at the end of Trainer initialization, and TRAINING_START should be called after Trainer.fit() is called. Note also that prepare_ddp_module() is now called within Trainer.__init__(), not Trainer.fit()

[ravi-mosaicml]

Surgery is fixed in #249, so we can now address this issue.

[ravi-mosaicml]

Addressed in #263. Kept Event.INIT; renamed Event.TRAINING_START to Event.FIT_START