Doing an allocation step before the limit is picked is more correct, but
it causes the idle scavenger and incremental marking to run more
frequently, which causes a regression on certain thread_time benchmarks.
This CL fixes the regression by reverting the behaviour. A proper fix
can be implemented in a follow-on.

Bug: chromium:795323
Change-Id: Ic62eab1ddfd1da4c5904e11eba7d5d71cee38f12
Reviewed-on: https://chromium-review.googlesource.com/835287
Commit-Queue: Ali Ijaz Sheikh <ofrobots@google.com>
Reviewed-by: Ulan Degenbaev <ulan@chromium.org>
Cr-Commit-Position: refs/heads/master@{#50239}