During instruction selection, match the pattern

  Float64LessThan(#0.0, Float64Abs(x))

which TurboFan generates for NumberToBoolean in the general case,
and which evaluates to false if x is 0, -0 or NaN, and compile this
to a simple (v)ucomisd using not_equal flags condition, which avoids
the costly Float64Abs.

Also consistently generate this pattern in TurboFan and the code
stubs/interpreter for ToBoolean.

R=mstarzinger@chromium.org
BUG=v8:5267

Review-Url: https://codereview.chromium.org/2473923003
Cr-Commit-Position: refs/heads/master@{#40740}