Add GRPO micro-batch size validation to prevent incorrect grouping and advantage computation

[aaghaazkhan]

Summary

This PR adds defensive validation to ensure GRPO's required grouping
invariants are not silently violated during rollout or log-probability
computation.

GRPO requires that for each prompt, the policy generates G completions
(num_generations). These G completions must stay grouped together
throughout rollout, reward computation, and log-prob computation.

If the configured micro-batch sizes are smaller than num_generations,
the framework silently reshapes rewards/logprobs incorrectly, mixing
generations from different prompts. This leads to invalid advantage
computations and results in the model failing to learn (e.g., 0%
accuracy for all metrics).

This PR introduces explicit validation checks to stop training early if
incorrect batching configurations are used.

---

What This Fix Does

Adds two safety checks:

1. Validate that rollout micro-batch size is large enough:

if self._rollout_micro_batch_size < self.algo_config.num_generations:
    raise ValueError(...)

2. Validate that log-prob computation micro-batch size is also large enough:

if self._compute_logps_micro_batch_size < self.algo_config.num_generations:
    raise ValueError(...)

These checks prevent GRPO from running with invalid batching
configuration that silently corrupts reward grouping.

Why This Matters

Without these checks:

• Batches are split incorrectly
• Rewards are grouped incorrectly
• Advantages are computed with wrong (prompt, generation) alignment
• Training converges to 0 accuracy even for simple tasks
• Debugging is very difficult because failure is something very unexpected

With this fix, the error is explicit, descriptive, and actionable.

---

Checklist:

• I have added all the necessary unit tests for my change.
• I have verified that my change does not break existing code and all unit tests pass.
• I have added all appropriate doc-strings/documentation.
• My PR is based on the latest changes of the main branch (if unsure, rebase the code).
• I have signed the Contributor License Agreement.
• I have followed Contribution Guidelines.