Allow offset indices in some linear algebra

[mcabbott]

This wants to make A * B work with OffsetArrays with minimal changes. The easy part is always to call similar(..., axes(A), ...) and never size.

The messier part is that it adapts the fallback generic_matmatmul! to allow offsets. The plan would be never to hit that fallback, by adding a method generic_matmatmul!(C::OffsetArray, ...) which removes offsets, and then calls mul!. That will allow an offset adjoint strided array (etc) will dispatch to the right place, after the offset is removed, without duplication or ambiguities. So perhaps the fallback need not exist?

Third, if we want to un-wrap the OffsetArrays and call mul! again, then we need α,β not MulAdd(α,β). I thought that the extra stage (and instability?) of making that twice was causing an extra allocation, so I added a step to call generic_matmatmul! without it, first. This does not seem to remove the allocaiton but might be the right thing to do anyway.

Finally, it might also be worth making generic_matmatmul! accept functions like adjoint instead of characters 'C'. In many cases this information is in the type domain when it is called, but α,β are not. (Last commit which does this unfortunately adds an ambiguity, right now.)