turns out LayerNorm also has weight and bias and needs to be pre-multiplied and trained for hypernets

Training UI Changes

Currently, trying to perform CLIP interrogation on a CPU fails, saying:

```
RuntimeError: "slow_conv2d_cpu" not implemented for 'Half'
```

This merge request fixes this issue by detecting whether the target
device is CPU and, if so, force-enabling `--no-half` and passing
`device="cpu"` to `clip.load()` (which then does some extra tricks to
ensure it works correctly on CPU).

Fix skip and interrupt buttons for Highres. fix option enabled

Add features to insert activation functions to hypernetworks