add g++ compiler to base image and implement hardware-aware FeedForward sharding

[prishajain1]

This PR introduces setup changes and optimizations needed to fix WAN trillium regression tests:

• Pre-baked g++ into the base system dependencies list. This provides the C++ compiler required by PyTorch Inductor (torch.compile) to compile the T5 text encoder at runtime, resolving the InvalidCxxCompiler error.
• Hardware-Aware Dynamic Sharding : Implemented dynamic, hardware-aware sharding inside NNXSimpleFeedForward:
  • Default (including Trillium/v6e): Shards embed to distribute memory and prevent OOM during high-resolution passes.
  • Ironwood (v7x): Keeps the embedding dimension replicated (None) to optimize cross-device communication bandwidth.

[github-actions]

e2e testgrid: https://8bcf50593faf4ea38060e236169827e5-dot-us-central1.composer.googleusercontent.com/dags/maxdiffusion_tpu_e2e/grid

[Perseus14]

Can you also add the python package av? This is required for ltx2 model post processing