Large fluctuations in eval results at step 0

[Bat-Reality]

Issue Description
When starting training with the current configuration, the evaluation results at step 0 show significant fluctuations. This makes the initial evaluation unreliable and affects the reproducibility of experiments.

Proposed Solution
Enable validation before training starts by setting the following in train_XXX.yaml:

trainer:
  val_before_train: True

Experimental Results
After applying the fix, evaluation results on gsm8k-eval at step 0 became much more stable (see figure below):

[pan-x-c]

val_before_train should not take effect. The above phenomenon is most likely caused by randomness.

In Trinity-RFT, evaluation is run in the Explorer module and we have removed the relevant components in the verl trainer.