Fixes #3802 (and closes its duplicate #3803, which was already closed manually).

Bug

After the TE 2.13 bump (#3800) the vision-encoder CUDA graph unit tests started failing with:

megatron/core/num_microbatches_calculator.py:19: AttributeError
E       AttributeError: 'NoneType' object has no attribute 'get'

The call chain is:

TestVisionTECudaGraphHelper.test_create_and_delete_cudagraphs
→ VisionTECudaGraphHelper.create_cudagraphs()
→ TECudaGraphHelper.create_cudagraphs()
→ _get_cuda_graph_input_data()
→ get_make_graphed_callables_kwargs() (defined inline)
→ get_num_microbatches() (megatron/core/transformer/cuda_graphs.py:2255)

get_num_microbatches() does _GLOBAL_NUM_MICROBATCHES_CALCULATOR.get(), but the global calculator was never initialized in this unit test (only the model-parallel state is initialized in setup_method), so the global is None.

The follow-up ci: Skip more tests in test_vision_cuda_graphs for LTS (#3860) merely marked the two tests @pytest.mark.flaky / @pytest.mark.flaky_in_dev, masking the failure rather than fixing it. The PP2 variant tracked under #3804 is a separate hang and is left untouched here.

Fix

In tests/unit_tests/transformer/test_vision_cuda_graphs.py:

1. Initialize the global num_microbatches calculator inside TestVisionTECudaGraphHelper._make_helper with the requested num_microbatches, matching the canonical pattern used in tests/unit_tests/transformer/test_cuda_graphs.py.
2. Destroy the calculator in teardown_method (and on each new _make_helper call) so the tests remain hermetic.
3. Drop the @pytest.mark.flaky / @pytest.mark.flaky_in_dev markers from test_create_and_delete_cudagraphs and test_create_cudagraphs_multi_microbatch so CI exercises the real code path again.

Minimal example of the fix shape:

def _make_helper(self, num_microbatches=1):
    destroy_num_microbatches_calculator()
    init_num_microbatches_calculator(
        rank=0,
        global_batch_size=self.micro_batch_size * num_microbatches,
        micro_batch_size=self.micro_batch_size,
        data_parallel_size=1,
        decrease_batch_size_if_needed=False,
    )
    return VisionTECudaGraphHelper(...)

Scope

PR scope is intentionally limited to the PP=1 TestVisionTECudaGraphHelper. The TestVisionTECudaGraphHelperPP2 variant has the same uninitialized-calculator latent issue but also a separate hang tracked in #3804; it will be addressed when #3804 is fixed.