Add Llama4VisionModel for multimodal decoding

[hengtaoguo]

Description

This PR implements a complete class Llama4VisionModel, by integrating all Llama4 basic vision components. It allows Llama4 multimodal decode by describing an input image. Joint by @hengtaoguo and @aireenmei .

• Core change Llama4VisionModel, which converts image tiles (batch_size, num_tiles, C, H, W) to feature activations (batch_size, num_tiles, num_patches, vision_output_dim_for_vit). Then Llama4MultiModalProjector projects it to (batch_size, num_tiles, num_patches, base_emb_dim). Example: (8, 5, 3, 336, 336) -> (8, 5, 144, 4096) -> (8, 5, 144, 5120). After that, the image tokens will be merged into text tokens by re-using merge_mm_embeddings.
• Refactor by adding a get_dummy_image_shape_for_init() to create desired dummy images for different models, for jit init purpose.

Tests

Tested full multimodal decode on v5p-16 cluster with this command, and workload with screenshot.

Checklist

Before submitting this PR, please make sure (put X in square brackets):

• I have performed a self-review of my code.
• I have necessary comments in my code, particularly in hard-to-understand areas.
• I have run end-to-end tests tests and provided workload links above if applicable.
• I have made or will make corresponding changes to the doc if needed.

[aireenmei]

Thanks Hengtao, great work!

[gagika]

thanks