Refactor shift operand handling for use from parse_operands, and defer
more constant handling to md_apply_fix3.  Unfortunately, this means
that we can no longer test the out-of-range diagnostic for that
constant in the same file as other diagnostics, because if the first
assembly pass produces errors, we don't bother running fixups.  This
isn't the only occurence of this problem in this patch series, and I
have not yet written replacement test cases for them.

It may not be obvious why the fields of the operands structure need to
be unsigned - it's so that they are not sign-extended, and therefore
do not clobber inst.instruction when or'ed in.  This hasn't been a
problem before because the values being stored in all prior operand
fields didn't ever get big enough to appear to be negative.
shift_kind, however, does.

zw