NCG: Really avoid the need for memory-to-memory moves in the register allocator

NCG: Really avoid the need for memory-to-memory moves in the register allocator

This is a follow-up to "NCG: Handle loops in register allocator".
The newly-introduced invariant that every virtual register is always assigned to
the same spill slot wasn't kept under all circumstances.
*Now* memory-to-memory moves should never be required when compiling hand-written
cmm code.