None of the crash checks were hit, so this seems to imply that
we're always dealing with non-nan values for the virtual index.

I'm not sure what's gone awry, but have simplified the while
loop here (and retested locally on ARM, x86, performance is
actually improved on modern ARM devices and slightly negative
on pre-64bit ARM devices). Results from media_perftests for
AudioConverterPerfTest.ConvertBenchmark

Z620 (X86): ~35% faster.
N5X (ARM64): ~3.7% faster.
N4 (ARM): ~0.81% slower.

BUG=663814
TEST=none

Review-Url: https://codereview.chromium.org/2552203003
Cr-Commit-Position: refs/heads/master@{#436852}