Pass dtype through to Gemma init for proper torch_dtype support

Init Gemma in float32, users control dtype via torch_dtype param

Remove Gemma dtype override

Uniform float32 weights, clean dtype handling: modeling_slip.py

Fix mixed-dtype handling in modeling_slip.py

Use set_default_dtype for uniform dtype init, assert batch size match in CrossAttention

Update modeling_slip.py: use HF torch_dtype instead of hardcoded bfloat16

Fix CrossAttention batch size mismatch: expand context to match query batch

Fix dtype mismatches: cast entire model to bfloat16 in init

Upload SLIP model, checkpoints, and source code

Upload SLIP model, checkpoints, and source code