Pixel classifier can fail if output is not opened in ImageJ

[alanocallaghan]

Steps to reproduce:

1. Open bioimageio extension dialog
2. Select a model that works (eg, affable-shark https://bioimage.io/#/?id=10.5281%2Fzenodo.5764892)
3. Click through dialog (with large downsample if you want it to be fast)
4. When prompted to open output in imageJ, select no
5. Select relevant pixel classifier in pixel classifier menu
6. Observe, an un-initialised field!

Affected versions:

• current main with qupath 0.6.0-rc (main)
• stable release jar with qupath 0.5.1

Ends up failing to enter the synchronized block in DjlDnnModel::predict because predictor is null. Suspect just a race condition

[alanocallaghan]

I can confirm this is reproducible across all classifiers, and I can't narrow it down to a bioimageio-specific issue. The predictor of the model of the DnnOp is null, even though the original model isn't lazy initilaised.