* Update common.py speed improvements

Eliminate .to() ops where possible for reduced data transfer overhead. Primarily affects warmup and PyTorch Hub inference.

* Updates

* Updates

* Update detect.py

* Update val.py