Fix RL LR schedule default

[AntonyMei]

RL computes its real training length from num_batches, rl.num_iterations, train_fraction, and num_epoch, but rl.yml inherited base.yml steps=150001. Config normalization then converted learning_rate_schedule_steps=-1 into that inherited step count before the RL optimizer saw max_train_steps, causing warmup/decay to use a much longer horizon than the actual RL run.

Set both steps and learning_rate_schedule_steps to -1 in the RL base config so the optimizer falls back to the computed max_train_steps unless a recipe explicitly pins the schedule horizon. Also warn when a positive learning_rate_schedule_steps differs from the computed RL training length, since that may be intentional but is easy to inherit accidentally.

Checklist

Before submitting this PR, please make sure (put X in square brackets):

• I have performed a self-review of my code. For an optional AI review, add the gemini-review label.
• I have necessary comments in my code, particularly in hard-to-understand areas.
• I have run end-to-end tests tests and provided workload links above if applicable.
• I have made or will make corresponding changes to the doc if needed, including adding new documentation pages to the relevant Table of Contents (toctree directive) as explained in our documentation.

[codecov]

Codecov Report

✅ All modified and coverable lines are covered by tests.

📢 Thoughts on this report? Let us know!