Only do shape detection instead of instance segmentation

[EricPedley]

Summary

Before:

After:

This lowers the runtime by like 20% but makes the accuracy worse. IMO we shouldn't actually merge this as it stands but every time we do data gen it might be worth training detection in case it works. Or, we try v8s instead of v8n to maybe undo the speed bonus but get better performance?

[EricPedley]

After merging with main again:

And results for the current main:

So it's slightly better :)

Gonna merge this in now. I guess the tile overlap thing really helped this.