main: L0_Unit_Tests_Megatron_Policy_1 fails, blocking functional tests on all PRs

[zhangyimi]

This unit test has been failing on main:

FAILED test_megatron_split_state.py::TestTrainMicrobatch::
       test_forwards_model_owned_packing_flags[model-owned-cp-slicing]

A guard in train_microbatch contradicts the test's parameters.

megatron_policy_worker.py:1465:

if self.media_placeholder_token_id is not None or self.model_slices_context_parallel_inputs:
    raise NotImplementedError("train_microbatch does not support multimodal models: ...")

test_megatron_split_state.py:446:

pytest.param(False, False, True, id="model-owned-cp-slicing")
#                         ^^^^ model_slices_context_parallel_inputs=True

The test sets model_slices_context_parallel_inputs to True and asserts it is forwarded, but the guard raises as soon as it sees that flag set. The case cannot pass.

The guard was added by #2957 (da39e1785); the parameter case predates it.

Once the unit tests fail, CI skips the functional stage, so no PR based on main gets functional coverage. Our PR #3566 is blocked on this.

Could a small fix land, or the case be marked xfail for now?

@ZhiyuLi-Nvidia