[Misc][LLaMa4] Compile LLaMa Vision Encoder layers

[Lucaskabela]

Purpose

We want to speedup up inference for mllama4 by applying torch.compile to the intensive workload, similar to what is done in #23207. We start by experimenting with the VisionEncoderLayer + PixelShuffle

Test Plan

vllm serve meta-llama/Llama-4-Scout-17B-16E-Instruct --tensor-parallel-size=8 --gpu_memory_utilization=.8 --max_model_len=8192

vllm bench serve   --backend openai-chat   --model meta-llama/Llama-4-Scout-17B-16E-Instruct    --endpoint /v1/chat/completions   --dataset-name hf   --dataset-path lmarena-ai/VisionArena-Chat   --hf-split train   --num-prompts 1000

Test Result

	Baseline (main)	This Pr
Successful requests	998	998
Benchmark duration (s)	998	998
Successful requests	72.52	62.15
Total generated tokens	117376	117504
Request throughput (req/s)	13.76	16.06
Output token throughput (tok/s)	1618.52	1890.73
Mean TTFT (ms)	35483.34	28623.5
Mean TPOT (ms)	264.74	233.7
Mean ITL (ms)	256.56	227.07

---

Essential Elements of an Effective PR Description Checklist

• The purpose of the PR, such as "Fix some issue (link existing issues this PR will resolve)".
• The test plan, such as providing test command.
• The test results, such as pasting the results comparison before and after, or e2e results
• (Optional) The necessary documentation update, such as updating supported_models.md and examples for a new model.
• (Optional) Release notes update. If your change is user facing, please update the release notes draft in the Google Doc.

[Lucaskabela]

<Updated to use dynamic dims - seems we still get very good speedup here!>

[Lucaskabela]

cc @zou3519 @ProExpertProg @ywang96