The current implementation overapproximates the
possible_nondeterminism_ bit by setting it whenever a NaN value is
reinterpreted as integer, or stored to memory. This hides bugs in the
interpreter that are handled as possible nondeterminism even though
they are not.
This CL fixes this by only setting the bit if a binary floating point
operation is executed and one of the inputs is a NaN.

R=ahaas@chromium.org

Bug: v8:6954
Change-Id: Ib937ae7730dbb140c012d07fae23b40ae7ed3d6b
Reviewed-on: https://chromium-review.googlesource.com/735599
Commit-Queue: Clemens Hammacher <clemensh@chromium.org>
Reviewed-by: Andreas Haas <ahaas@chromium.org>
Cr-Commit-Position: refs/heads/master@{#48954}