[ktf]

invokeProcess walked all members of the task seven times per process
function to bind partitions and grouped combinations, instantiating the
binding lambdas for every member, although only those two kinds act on
them; for everything else the calls are empty overloads. Select the
partitions and combination generators once per task type, with a named
selector so that the full member walk is shared by all process
functions, and run the binding lambdas over those only. The calls that
do something happen in the same order as before.

StructToTuple gains apply_refs_tuple_cat_sized, a member walk whose
callable returns a tuple per member, concatenated into one.

Full -O2 compile, Apple clang 21, before -> after:

• femtoUniverseProducerTask: 66.5 s -> 58.4 s (macro path, as for
  GCC 14), 67.6 s -> 58.7 s (structured binding packs)
• dataCreatorCharmResoToDstarReduced: 44.6 s -> 41.7 s
• derivedDataWriter: 79.9 s -> 74.8 s
  In femtoUniverseProducerTask the instantiation of the member walks goes
  from 5.9 s to 1.4 s.

[ktf]

@aalkin apparently selecting the bindables upfront avoids instanciating the large types at each invokation (for some 5% speedup while building O2Physics).

[aalkin]

This is similar to what we have seen with the other pack operations, since it is always combinatorics. For iterating on the preselected tasks members and methods we need C++26 to properly fix this.