[tinkerification] Remove scheduler from inside the workers and default to use set_lr() and get_lr() in training loop

[erictang000]

After #978, we should remove the scheduler logic in SkyRL-Train workers and just rely on the tinker convention of setting learning rate for each optim_step() call

cc: @tyler-griggs

[jonahsamost]

I can take this