Add lr_warmup_steps, max_steps values validation

[shenxiangzhuang]

Fix #1393

[rasbt]

Thanks for contributing! This looks good to me overall. Let me take care of the tests.

[rasbt]

Actually, the tests are failing because in the README test, I am using

litgpt finetune_lora" checkpoint_dir \
--train.max_steps 1" \
...

With this PR, we will not be able to run quick test runs without also having to specify the warm-up steps via an appropriate value (the current default is 100).

I'd say maybe this should not be an error but perhaps a warning instead. What do you think @awaelchli & @Andrei-Aksionov ?

[awaelchli]

I think the suggestion for a warning makes sense.

[rasbt]

Thanks again for the contribution @shenxiangzhuang, these user experience improvements are well appreciated!