Make the WarmupCosineLR ratio flags reach the config

[vineethsaivs]

add_tuning_arguments declares --warmup_min_ratio and --cos_min_ratio for
WarmupCosineLR, but get_config_from_args routed every schedule that is not
LRRangeTest or OneCycle through override_warmupLR_params. So both flags
were parsed and then dropped, and the config came back carrying the WarmupLR
parameters instead:

$ parser = add_tuning_arguments(argparse.ArgumentParser())
$ args = parser.parse_args(["--lr_schedule", "WarmupCosineLR",
                            "--warmup_min_ratio", "0.1", "--cos_min_ratio", "0.05"])
$ get_config_from_args(args)
({'type': 'WarmupCosineLR',
  'params': {'warmup_min_lr': 0, 'warmup_max_lr': 0.001,
             'warmup_num_steps': 1000, 'warmup_type': 'log'}}, None)

Both ratios are gone, and building the scheduler from what is left fails,
because WarmupCosineLR has no min/max lr at all:

TypeError: WarmupCosineLR.__init__() got an unexpected keyword argument 'warmup_min_lr'

WARMUP_MIN_RATIO and COS_MIN_RATIO were already defined next to the other
config keys, so only the plumbing was missing.

get_lr_from_config had the mirror-image problem: it falls through to
lr_params[WARMUP_MAX_LR] for anything that is not LRRangeTest or
OneCycle. For WarmupCosineLR that returned an unrelated WarmupLR default on
a CLI-built config, and raised KeyError: 'warmup_max_lr' on a config
actually written for this schedule. WarmupCosineLR scales each param group's
own lr by a ratio, so there is no learning rate in its params to read, and it
now says so instead.

The JSON config path was never affected: engine._configure_lr_scheduler
calls scheduler(optimizer, **scheduler_params) straight from the config
file, so "scheduler": {"type": "WarmupCosineLR", "params": {"warmup_min_ratio": 0.1}}
has always worked. Only the add_tuning_arguments CLI helper is fixed here.

Other schedules are untouched. LRRangeTest, OneCycle, WarmupLR and
WarmupDecayLR produce the same params and the same get_lr_from_config
answer as before, and test_other_schedules_keep_their_config_params pins
that.

Not included, so as to keep this to one thing: TOTAL_NUM_STEPS is
defined but unused, and neither WarmupDecayLR nor WarmupCosineLR can be
built from get_config_from_args output alone because both need
total_num_steps, which has no CLI flag. Callers pass it themselves today.
Happy to add that separately if you would like it plumbed.

Tests

tests/unit/runtime/test_lr_schedulers.py gains three plain (non-distributed)
tests next to the existing ones:

• test_warmup_cosine_lr_config_from_args_carries_the_ratios
• test_warmup_cosine_lr_has_no_lr_in_its_config_params
• test_other_schedules_keep_their_config_params

DeepSpeed does not build on this machine (macOS, no CUDA), so I ran the three
new test bodies against the unmodified lr_schedules.py from master and
against the patched one, loading the module directly:

before: 2 failed, 1 passed
        (KeyError: 'warmup_max_lr', and the ratios missing from the config)
after:  3 passed

The one that passes on both sides is the no-regression control.

yapf --style .style.yapf --diff is clean on both changed files, and neither
adds a line over the 119 column limit.

[chatgpt-codex-connector]

Codex usage limits have been reached for code reviews. Please check with the admins of this repo to increase the limits by adding credits.