Fix RCSafetyModule offline learning guard to actually raise

[rootkiller6788]

What

RCSafetyModuleCostCriticContinuousAction.learn_batch is meant to reject offline safety learning with a clear error, but the AssertionError(...) was never raised:

def learn_batch(self, batch: TransitionBatch, policy_learner: PolicyLearner) -> None:
    """Don't support offline training for now."""
    AssertionError("Offline safety learning is not supported yet.")
    pass

The bare expression constructs the exception and discards it. In addition, the signature (self, batch, policy_learner) mismatches both the base class SafetyModule.learn_batch(self, batch) and the only call site PearlAgent.learn_batch (self.safety_module.learn_batch(batch)), so a user hitting the offline path gets a TypeError: learn_batch() missing 1 required positional argument: 'policy_learner' instead of the intended message.

Why

Offline safety learning is documented as unsupported, so the method should fail fast with an informative error rather than silently no-op (or crash with an unrelated TypeError).

How

• Align the signature to (self, batch) (matches the base class, IdentitySafetyModule, RiskSensitiveSafetyModules, and the agent call site).
• raise the AssertionError, and drop the now-resolved # pyre-fixme[14].
• Add test/unit/with_pytorch/test_safety_module.py asserting learn_batch raises.

Verified: new test fails before the change (TypeError) and passes after; test_agent.py (agent construction) suite passes with the change.