Log learning_rate to wandb during training

[vertix]

Summary

• OpenPI never logs the learning rate, so it is invisible in wandb. This forces an external monkeypatch (the proposed positronic/vendors/openpi/_launch.py in Span OpenPI LR schedule across the full run via --lr-schedule.decay-steps positronic#391).
• Recreate the LR schedule in the training loop (it is otherwise hidden inside the optimizer) and log its value into the same wandb.log payload as the other metrics, at every log_interval.
• This makes the wandb half of _launch.py unnecessary. The other half (tying decay_steps to run length) needs no openpi change: lr_schedule is already CLI-exposed, so the caller can pass --lr-schedule.decay-steps=<num_train_steps>.

Test plan

• uvx ruff check scripts/train.py passes.
• Manual: run a short training job and confirm learning_rate appears in wandb with the expected warmup/cosine shape (cannot run here - training requires CUDA).