Remove LR scheduler for workers

[jonahsamost]

Referencing #998

Remove internal learning rate scheduler logic from the skyRL train workers.

Since #978 added set_lr() support, Tinker can now control the LR externally. Since the workers's LRs could conflict with Tinker's LR, we just remove the ability for workers to set their own LRs

[gemini-code-assist]

Code Review

This pull request effectively removes the internal learning rate scheduler logic from the training workers, centralizing LR control under an external system like Tinker. The changes are consistent and thorough across FSDP and Megatron strategies, abstract base classes, and worker implementations. Key modifications include removing scheduler instantiation, step() calls, and state management from checkpoints. The code is now cleaner and the responsibility for learning rate scheduling is clearly externalized. The changes look solid.

[gemini-code-assist]

With the removal of get_megatron_optimizer_param_scheduler from the imports, the function itself in skyrl_train/skyrl_train/distributed/megatron/optimizer.py appears to be dead code. Consider removing it in a follow-up to improve code maintainability.

[pcmoritz]

@erictang000 Do you think this makes sense to do / when would be the right time to merge it? As we fully standardize on the Tinker API, I feel like this should be removed, right?