Optimize weighting of standard & phase loss and / or stop metric

[alisandra]

in particular, I noticed that for vertebrates the phase f1 often improved with a strong delay compared to genic f1, i.e. on many less-than-lucky models the phase remained severely under trained at training end.

I simply took 'more lucky' models for now, but one should come back to it.

Might also be something in the data causing this, as vertebrates was the only group there was an eye-catching difference for.

[alisandra]

Phase for CDS start != 0 was flipped, this is fixed in dev now. Will need to add tests and merge to main before closing.