Improve memcopy loops in code generated by the CakeML compiler

#798 correctly points out that memcopy loops are sometimes suboptimal -- one such loop in the inner most loop of the GC!

This issue is about improving memcopy loops so that they generate better loops when possible. For register constrained targets, it might not be possible to unroll these memory copying loops in the heart of the GC, but for most targets there should be spare registers that can be used for this purpose.