Add better code for finding `SL(2,q)` inside `SL(4,q)`

[fingolfin]

We need this for the constructive SL(d,q) recognition.

We currently do this:

  if not (q in [2,3,4,5,9]) then
     return RECOG.SLn_godownfromd(h,q,4,d);
  else
     return RECOG.SLn_exceptionalgodown(h,q,d);
  fi;

Instead we should implement what is described in Daniel's thesis in section 5.1.3 resp. 5.1.4. A start of that is in RECOG.OneEvenSL4 in PR #371, but it is incomplete. Looking at the code, I note that it also defines a new function RECOG.InvolutionCentraliser -- which actually overwrites an existing (but unused RECOG.InvolutionCentraliser).

But I also wonder if we even need this: the algorithm uses random elements of the involution centralizer. And to compute the that centralizer, we generate... random elements of it, via a helper RECOG.ChFromg. So perhaps we can just directly use the output RECOG.ChFromg, which would avoid some overhead.

But perhaps instead use the existing function RECOG.CentralisingElementOfInvolution defined elsewhere?

[Till-Eisen]

I took a look at section 5.1.3 where the algorithm GoingDownFinalStepSL is described.

Regarding the choice of $\ell$ in line 11: Let $A (, D)$ be the upper (lower) $2\times 2$-block of $g$. The condition in line 11 requires $D^\ell = I$, so $\mathrm{ord}(D) \mid \ell$. To maximise the chance that the upper block $A^\ell$ remains non-trivial, one should take $\ell = \mathrm{ord}(D)$ as the minimal such choice. If $\mathrm{ord}(D) \mid \mathrm{ord}(A)$, then $A^\ell = I$ regardless and $g$ should be discarded..