fix: add embed_vision to MULTIMODAL_SUFFIXES and set lbs=2 for Gemma4 PP2 recipe

[khazic]

What does this PR do ?

Two small follow-up fixes to #1904, caught in review by @HuiyingLi: correct `embed_vision` stage-0 assignment for Gemma4 PP, and fix `local_batch_size` in the PP2 recipe to avoid a 100% pipeline bubble.

Changelog

• Add `embed_vision` to `MULTIMODAL_SUFFIXES` in `hf_utils.py` so Gemma4's vision projection layer is correctly assigned to PP Stage 0 alongside `vision_tower`. It was missing because original PP testing used pure-text inputs only.
• Set `local_batch_size: 2` in `gemma4_31b_tp4_pp2.yaml`. `lbs=1` technically runs (the per-rank stage count for 1F1B is 1, so no warning triggers), but produces zero pipeline overlap (100% bubble). Local runs were always done with `lbs=2`; the recipe was submitted with `lbs=1` by mistake.

Before your PR is "Ready for review"

Pre checks:

• Make sure you read and followed Contributor guidelines
• Did you write any new necessary tests?
• Did you add or update any necessary documentation?

Additional Information

• Related to feat: TP+PP support for Gemma4 VLM (with tied lm_head fix and 31B recipes) #1904

[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.

[akoumpa]

/ok to test e97b160

[HuiyingLi]

Thank you so much!