Speeding up `vp-setupfit`

Created by: giacomomagni

It looks to me that vp-setupfit is really slowed down by the fact that predictions are computed for all the replicas:

https://github.com/NNPDF/nnpdf/blob/ce6c05cacec95dee36c577503c9a1109daa2908a/validphys2/src/validphys/convolution.py#L132-L180

even though only the central prediction is then needed:

https://github.com/NNPDF/nnpdf/blob/ce6c05cacec95dee36c577503c9a1109daa2908a/validphys2/src/validphys/theorycovariance/construction.py#L179

Since by default set we are using 210715-n3fit-1000-001 which contains 1000 replicas, this could be a major wast of time.

Similarly as done for the t0pdfset here there is no need to load the entire pdf set: https://github.com/NNPDF/nnpdf/blob/ce6c05cacec95dee36c577503c9a1109daa2908a/validphys2/src/validphys/lhapdfset.py#L44-L52

Thanks to @andreab1997 and @RoyStegeman for the assistance