When I centralized the treatment of memory operands, I forgot to delete
the old code.

There is a semantic difference between the old and new code. The old
code was handling either memory operands, or ranges that had a spilled
predecessor. The new code handles just memory operands. It may
happen that (using LinearScan) an active range is spilled when trying
to allocate another range (see SplitAndSpillIntersecting). That may make
it a candidate for the old version of the code, however, since we would
have spilled up to a register use, the old code wouldn't have had taken
effect.

Perf data shows this nuance doesn't make a difference in perf.

BUG=

Review URL: https://codereview.chromium.org/1428943004

Cr-Commit-Position: refs/heads/master@{#31729}