[3outeille]

What does this PR do?

Stack: #50493 → #50496 → #50511 → #50514 → this PR → #50521 → #50522

Adds ContextParallelTesterMixin to CausalLMModelTest, so every causal LM gets two tests, test_cp_matches_unsharded and test_cp_matches_unsharded_packed. Each test runs on 2 CPU ranks with cp_size=2 and checks that the loss, the logits of each rank's slice, and every gradient match an unsharded run of the same weights.

• The tester configs run unchanged. The only change is output_router_logits=False, because the MoE auxiliary loss is not a sum over tokens.
• If CP refuses a model, the test checks that loading with cp_size=2 raises, then skips with the reason.
• An allow list (CONTEXT_PARALLEL_TEST_MODEL_TYPES) limits CI to 7 model types, like the FSDP mixin. The tests carry is_fsdp_test.

Tests

With the allow list off, over the 95 causal LM test classes:

result	classes
pass (plain and packed)	54
refused with a clear error	41
fail	0

The refusals: attention outside the attention interface, eager-only models, linear attention, recurrent or conv layers, sparse indexers, a single key/value head, layers the default plan does not name (GPT-2, jetmoe, ...), and the models that set base_model_cp_plan = None in #50496 (gemma3n, gemma4 and gemma4_unified, which inherits it, beam, ministral3).

[HuggingFaceDocBuilderDev]

The docs for this PR live here. All of your documentation changes will be reflected on that endpoint. The docs are available until 30 days after the last update.