Replace VideoMAE sinusoid encoding helper with PyTorch implementation

[praful-srinivasan-027]

What does this PR do?

This PR replaces the NumPy implementation of get_sinusoid_encoding_table in modeling_videomae.py with an equivalent vectorized PyTorch implementation.

As part of this refactor, fixed sinusoidal positional embeddings are registered as non-persistent buffers and reinitialized in _init_weights, making the model compatible with meta device initialization while preserving the existing sinusoidal encoding behavior.

This also resolves the existing TODO to implement the helper using PyTorch.

Current status: This PR is opened as a draft while investigating a model-parallel regression introduced by the buffer registration changes. The meta initialization tests now pass, but test_model_parallelism still requires investigation.

[github-actions]

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

run-slow: videomae

[praful-srinivasan-027]

Hi @yonigozlan! I opened this as a draft because I'd appreciate some guidance.

While replacing the NumPy implementation of get_sinusoid_encoding_table with a PyTorch version, I updated the fixed positional embeddings to be registered as non-persistent buffers and reinitialized them in _init_weights so that the model passes the meta initialization tests.

However, this introduces a failure in test_model_parallelism due to a device mismatch between sequence_output and the positional embeddings after model sharding.

Before I continue, I wanted to check whether this is the expected approach for making these fixed positional embeddings compatible with meta initialization. If so, is there a recommended pattern for handling the model-parallel case as well? If not, I'd appreciate any guidance on the preferred approach.

Thanks!

[github-actions]

CI Dashboard: View test results in Grafana