The failure of dx test 105 wasn't actually spurious. Fixed the code.

In particular, for array load and store instructions, dx was being too
aggressive about letting the contents of the stack override the
implied type expected by the instructions. This led to cases where dx
would emit code instead of reporting an error.

The failure on the array load side of things implied that there also
needed to be a new regression test on the store side, since the
existing array store test didn't catch this. So, I added that too in
this change.

Finally, I took this opportunity to do minor whitespace-type cleanups
in a couple of related files that I opened during the course of the
investigation.

Change-Id: I72c644f66afb1108ae43a129ac81b010d072155a