Use inference_mode instead of no_grad

[bryant1410]

Use inference_mode instead of no_grad for BaseHandler. It should bring a (small?) performance boost.

Though, I have 2 questions:

• What's the min supported PyTorch version? Because inference_mode is supported since PT 1.9.
• Should other handlers (including example ones) be changed as well?

[msaroufim]

A bit conflicted about this one since inference mode has a bad interaction with torch.compile

[bryant1410]

A bit conflicted about this one since inference mode has a bad interaction with torch.compile

Yeah, I've seen that. I think it's only for 2.0 but not 2.1: pytorch/pytorch#103132. It works fine for me for 2.1 (but not with 2.0).

When compile is being used, we could only use it for PT > 2.0, else use no_grad. What do you think?