Before this CL, we created one live range per successive set of
deferred blocks. For scenarios with many such blocks, this creates
an upfront pressure for the register allocator to deal with many ranges.
Linear sorts ranges, which is a super-linear operation.

The change places all deferred intervals into one range, meaning that,
at most, there will be twice as many live ranges as the original set. In
pathological cases (benchmarks/Compile/slow_nbody1.js), this change
halves the compilation time. We see some improvements elsewhere,
notably SQLite at ~4-5%.

We may be able to avoid the subsequent merge. Its cost is the
additional ranges it may need to create. The sole reason for the merge
phase is to provide an unchanged view of the world to the subsequent
phases. With the at-most-one splinter model, we may be able to teach
the other phases about splintering - should we find perf hindrances
due to merging.

Review URL: https://codereview.chromium.org/1391023007

Cr-Commit-Position: refs/heads/master@{#31224}