Add scale_warmup argument to schedulers

[hanlint]

Based on request from @codestar12 , this PR adds the option to also scale the warmup period for our schedulers with the scale_schedule_ratio.

For example:

    from composer.optim.scheduler import MultiStepWithWarmupScheduler

    scheduler = MultiStepWithWarmupScheduler(
        milestones=["10ep", "20ep"],
        t_warmup="4ep",
        scale_warmup=True,
    )

With scale_schedule_ratio=0.5, this scheduler will warmup for 2 epochs, then step at 5 and 10 epochs. scale_warmup defaults to False to preserve the current behavior.

During implementation, I observed an unintuitive default behavior with our warmup. Suppose we have max_duration=100ba. If we define a scheduler with a warmup in batches (e.g. t_warmup=10ba), and apply an ssr, the warmup period will not be scaled, per our default behavior. However, if the warmup was defined in duration (e.g. t_warmup=0.1dur), and apply an ssr, then the warmup period will always be scaled (as we scale max_duration in the trainer beforehand). This PR respects that behavior, but perhaps we should attempt a fix separately.

Closes https://mosaicml.atlassian.net/browse/CO-668

[ravi-mosaicml]

CC @jbloxham since he originally wrote the schedulers
CC @jfrankle for whether schedulers should be scaled for warmup -- I was under the impression that they shouldn't, but I don't know the research background here. Though I suppose if it's a non-default option, then that's fine.

[ravi-mosaicml]

LGTM. Can you create a JIRA to track the fix for scale_warmup=False when the warmup period is specified in dur?