In load_pretrained helper: Fail if cannot find pretrained model

[dnouri]

Previously, this would only log a warning.

Motivation

Calling e.g. timm.create_model('mobilenetv3_small_075', pretrained=True) to retrieve a pre-trained model will sometimes fail silently (or: only with a warning), potentially leading to hard-to-debug cases where you use a supposedly pretrained model downstream and get unexpected results. (Consider that I was using Git master while my colleague used the latest timm release; the call to create_model didn't fail for both of us, yet my colleague's version behaved differently for less than obvious reasons.)

I believe it would be much less surprising if the call to timm.create_model(name, pretrained=True) actually failed in case a pretrained version wasn't available. Retrieving a randomly initialized version instead is I think not what the user wants when they explicitly set pretrained=True.