fix failed unit test for glm_moe_lite model

[kaixuanliu]

This PR fixes following failed test cases:

FAILED tests/models/glm4_moe_lite/test_modeling_glm4_moe_lite.py::Glm4MoeModelTest::test_eager_matches_fa2_generate
 - RuntimeError: mat1 and mat2 shapes cannot be multiplied (14x256 and 512x32)
FAILED tests/models/glm4_moe_lite/test_modeling_glm4_moe_lite.py::Glm4MoeModelTest::test_flash_attention_2_continue
_generate_with_position_ids - RuntimeError: mat1 and mat2 shapes cannot be multiplied (91x256 and 512x32)
FAILED tests/models/glm4_moe_lite/test_modeling_glm4_moe_lite.py::Glm4MoeModelTest::test_flash_attention_2_padding_
matches_padding_free_with_position_ids - RuntimeError: mat1 and mat2 shapes cannot be multiplied (91x256 and 512x32
)
FAILED tests/models/glm4_moe_lite/test_modeling_glm4_moe_lite.py::Glm4MoeModelTest::test_flash_attention_2_padding_matches_padding_free_with_position_ids_and_fa_kwargs - RuntimeError: mat1 and mat2 shapes cannot be multiplied (91$256 and 512x32)
FAILED tests/models/glm4_moe_lite/test_modeling_glm4_moe_lite.py::Glm4MoeModelTest::test_flash_attn_2_equivalence $ RuntimeError: mat1 and mat2 shapes cannot be multiplied (91x256 and 512x32)
FAILED tests/models/glm4_moe_lite/test_modeling_glm4_moe_lite.py::Glm4MoeModelTest::test_flash_attn_2_fp32_ln - Ru$timeError: mat1 and mat2 shapes cannot be multiplied (91x256 and 512x32)
FAILED tests/models/glm4_moe_lite/test_modeling_glm4_moe_lite.py::Glm4MoeModelTest::test_flash_attn_2_from_config $ RuntimeError: mat1 and mat2 shapes cannot be multiplied (91x256 and 512x32)
FAILED tests/models/glm4_moe_lite/test_modeling_glm4_moe_lite.py::Glm4MoeModelTest::test_flash_attn_2_inference_eq$ivalence - RuntimeError: mat1 and mat2 shapes cannot be multiplied (7x256 and 512x32)
FAILED tests/models/glm4_moe_lite/test_modeling_glm4_moe_lite.py::Glm4MoeModelTest::test_flash_attn_2_inference_eq$ivalence_right_padding - RuntimeError: mat1 and mat2 shapes cannot be multiplied (7x256 and 512x32)
FAILED tests/models/glm4_moe_lite/test_modeling_glm4_moe_lite.py::Glm4MoeModelTest::test_flash_attn_kernels_infere$ce_equivalence - RuntimeError: mat1 and mat2 shapes cannot be multiplied (7x256 and 512x32)

This is because the test configuration used the default value of 256 for v_head_dim, which caused a dimension mismatch when padding and slicing in flash attention. After this PR all above cases can get passed.

[github-actions]

[For maintainers] Suggested jobs to run (before merge)

run-slow: glm4_moe_lite