Lora for remaining transformer models to inference_exp

[hansent]

Description

This add inference_exp support for remaining transformer models that we can fine tune with LoRA on roboflow.

• qwen model implementation added to inference_exp
• paligemma: added lora/peft support and fixed preprocessing
• smolvlm: added lora/peft support and fixed preprocessing, changed to pass tensors instead of converting to PIL images

for all three models:

• each is registered on staging and prod model registry as base model and one lora test model packages are:
• registered base models: paligemma2-3b-pt-224, qwen25vl-7b, smolvlm-256m, smolvlm (regular 1.7B size)
• registered lora test models: paligemma-lora-test, qwen-lora-test, smolvlm-lora-test
• e2e tests that load based and lora version model from model registry and run predictions
• model prediction tests
• preprocessing tests

Type of change

• Bug fix (non-breaking change which fixes an issue)
• New feature (non-breaking change which adds functionality)
• This change requires a documentation update

How has this change been tested, please provide a testcase or example of how you tested the change?

integration tests locally

Any specific deployment considerations

n/a

Docs

n/a

[PawelPeczek-Roboflow]

should be smolvlm-v2 probably

[hansent]

changed