Pull Request: Bug Fixes for CUDA Custom Ops and timm Layer Import

[moqingyan]

Summary
This pull request addresses two main issues:

1. Deprecation warning from timm: Updated imports from [timm.models.layers]to [timm.layers] to resolve the deprecation warning.
2. CUDA custom C++ ops loading: Update value.type() to value.scalar_type() for torch support.

Testing
from groundingdino.util.inference import Model

Without the fix, we have warnings:

FutureWarning: Importing from timm.models.layers is deprecated, please import via timm.layers
UserWarning: Failed to load custom C++ ops. Running on CPU mode Only!