Use SubstitutePID rather than ParticleCombiner for making combinations for SLB isolation tool

There was a flaw in our methodology for applying the iso MVA before, the old methodology was this:

1. Find all child particles from the ground-state line (e.g. in B\to D\tau, this would be the D and the \tau
2. Reconstruct those into a flightless parent (typically a B^*)
3. Combine that parent with all other pions in the decay and evaluate the MVA on those combinations

This is flawed for the case of multiple candidates. For example, if step 1 finds two decays, that's two D's and two \tau's which will be made into 4 parents in step 2.

The new methodology is therefore:

1. Find the parent(s) from the line
2. convert the species of that parent to be flightless (again a B^*)
3. as above

This has the added benefit of a higher throughput (sprucing throughput appears to increase globally by about 12%)

Now it appears that fewer combinations are made, however, this is just due to the false combinations being excluded. I attach the overlaid plots showing the MVA input and output distributions before and after the fix (red for before, blue for after).

Since the distributions are very heavily background dominated, and what we have removed is also background-dominated. It appears in these plots as just a scaling-down across the board.

Furthermore, we believe that the method we used to generate our tuples for the training didn't have this bug - and thus we do not need to retrain the models.

Replaces !3292 (closed)

This also integrates !3696 (closed), which fixes a bug in the calculation of Delta R

MR is dependent on: lhcb-datapkg/ParamFiles!106 (merged)

@amathad