Fix performance regression in AllHomomorphismClasses and possibly other code searching isomorphisms

[fingolfin]

Since this is costly, it is worth to force a minimal generating set for
solvable groups. This resolves problems the yangbaxter package has had.

Close #5024

This patch is taken from @hulpke's PR #5031, but having different, unrelated improvements and bug fixes on separate PRs reduces the work needed to make GAP releases quite a bit (as it allows us to deal with those PRs in automation, instead of adding them to the list of PRs that need manual work), and is also more convenient for reviewers, so I am splitting it off here.

For reference, here is the motivation for this patch from the description of PR #5024:

The early abort condition Length(gens)<=Length(AbelianInvariants(G))+2
was inserted in 9d380c4 to speed up
calculations in certain large cases. However this caused regressions
in various package test suites, so bdf94cb
attempted to address this by reordering some computations. Alas, that
was not enough: the YangBaxter test suite got slowed down by an order of
magnitude (it went from about 30 seconds to several minutes on my laptop).

This PR addresses this.

[fingolfin]

Backported to stable-4.12 via fa8d50a