Regression: `Matrix(GF(3), [[]]);` now raises an error

[fingolfin]

This has been breaking the semigroups tests (CC @james-d-mitchell) in the package distro since yesterday:

gap> Matrix(GF(3), [[]]);
Error, cannot convert <m> to 'Is8BitMatrixRep'
*[1] Error( "cannot convert <m> to 'Is8BitMatrixRep'" );
   @ /Users/mhorn/Projekte/GAP/gap/lib/vec8bit.gi:1135
 [2] method( req1, req2, req3, req4 )
   @ /Users/mhorn/Projekte/GAP/gap/lib/methsel.g:254
 [3] NewMatrix( DefaultMatrixRepForBaseDomain( R ), R, l, list )
   @ /Users/mhorn/Projekte/GAP/gap/lib/matobj.gi:385
<function "Matrix method">( <arguments> )
 called from read-eval loop at *stdin*:1
you can 'quit;' to quit to outer loop, or
you can 'return;' to continue

I have not yet checked what causes it but of course PR #6369 is high up on the suspect list (CC @ThomasBreuer)

[ThomasBreuer]

The Matrix method calls NewMatrix( DefaultMatrixRepForBaseDomain( R ), R, l, list ) where DefaultMatrixRepForBaseDomain( R ) is Is8BitMatrixRep.
In earlier versions, the NewMatrix method simply called ConvertToMatrixRep without checking whether the conversion worked. Now the result of ConvertToMatrixRep is checked, here it is fail because Is8BitMatrixRep does not support matrices with zero columns.

The logical problem seems to be that DefaultMatrixRepForBaseDomain depends only on R but there are representations for which not all NewMatrix calls will be possible.
I think NewMatrix is right not to return an object in a representation different from the requested one.
Should NewMatrix be allowed to return fail in such situations, and then Matrix can choose a suitable representation?
(If yes then can we generically decide to choose IsPlistRep?)

[ThomasBreuer]

One more issue:

gap> m:= Matrix(GF(3), [], 0 );
[  ]
gap> IsPlistRep( m );
true

The result is not a matrix.

Here both Matrix and NewMatrix do not run into errors because ConvertToMatrixRep is cheating.

gap> m:= [];
[  ]
gap> ConvertToMatrixRep( m, 3 );
3
gap> m;
[  ]
gap> IsPlistRep( m );
true

[fingolfin]

So the call used to work but strictly speaking the result was incorrect.

We currently can't really return a "correct" result (in filter Is8BitMatrixRep), because we can't really do compressed matrices with zero rows (that would require changing the object layout to provide space for the number of columns -- something I am interested in doing, but it is not yet here), no matrices with zero columns (that would require empty compressed vectors; those are possible but still risky, see also #2121).

@james-d-mitchell Semigroups has this test; is this crucial, or could it be removed there? By that I mean: presumably you have more serious functionality that might require this to "work"? Although the Semigroups test suite does not seem to hit upon any other situations...

# Test Matrix for a finite field and list consisting of an empty list
gap> Matrix(GF(3), [[]]);
[ [  ] ]