Add 3D video patch embedding for Diffusion Transformers

[CodersAcademy006]

This PR adds a 3D patch embedding module for video inputs, enabling
Diffusion Transformer (DiT)–style video models in Megatron-LM.

The module converts video tensors [B, C, T, H, W] into a sequence of
tokens [B, N, D] using a single Conv3D projection, consistent with ViT
and DiT architectures.

• Introduces VideoPatchEmbed under megatron/core/vision
• Uses Conv3D for efficient linear projection over spatiotemporal patches
• Produces Transformer-ready token sequences
• Includes unit test validating shape correctness
• No impact on existing models or training paths

This PR is intentionally minimal and self-contained.

Fixes #2796 - Part - 3 DiT reference wiring

[copy-pr-bot]

This pull request requires additional validation before any workflows can run on NVIDIA's runners.

Pull request vetters can view their responsibilities here.

Contributors can view more details about this message here.