Fix empty LR scheduler for old jobs, clean up argument validation

[mryab]

The introduction of Literal in #273 broke parsing of old jobs without the learning rate scheduler specified, as Pydantic receives an empty string for the type and can't match it to any of the LR scheduler types that we have defined.

The error looks like this:

data.4.lr_scheduler.FinetuneCosineLRScheduler.lr_scheduler_type
  Input should be 'cosine' [type=literal_error, input_value='', input_type=str]
    For further information visit https://errors.pydantic.dev/2.10/v/literal_error

As a simple fix before any backend changes, this PR adds an empty scheduler type and enables it only in FinetuneResponse. together fine-tuning list works with the new fix

The PR also makes a few cleanups around the codebase related to LR scheduling and fixes a hidden bug introduced in the previous PR, where lr_scheduler_args was renamed to lr_scheduler but Pydantic silently accepted the previous name of the field

[mryab]

As you can see here, we passed lr_scheduler_args despite this field not being defined for FinetuneCosineLRScheduler

[azahed98]

Oops ty for the catch. I'm actually not sure it didn't have a validation error before...

[thepowerfuldeez]

Smart! Looks good to me if we have no other options to support old jobs