So far, we've been moving down gaps wholesale. This change moves
individual move operations instead. This improves some benchmarks,
and should overall reduce code size, because it improves the chance of
reducing the number of moves.

For example, there are improvements on x64 in Emscripten (Bullet, in
particular) , JetStream geomean, Embenchen (zlib).

In the process of making this change, I noticed we can separate the
tasks performed by the move optimizer, as follows:

- group gaps into 1
- push gaps down, jumping instructions (these 2 were together before)
- merge blocks (and then push gaps down)
- finalize

We can do without a finalization list. This avoids duplicating storage -
we already have the list of instructions; it also simplifies the logic, since,
with this change, we may process an instruction's gap twice.

Compile time doesn't regress much (see pathological cases), but we
may want to avoid the allocations of the few sets used in the new code.
I'll do that in a subsequent change.

BUG=

Review URL: https://codereview.chromium.org/1634093002

Cr-Commit-Position: refs/heads/master@{#33715}