For NumberAdd/Subtract/Multiply we currently onlt consult the upper
bound to decide whether to compute using Int32 or Float64 operations,
whereas for NumberModulus, NumberEqual, etc. we do decide based on
the feedback types, where the only significant difference is that we
cannot promise Word32 truncations on the inputs.

This change unifies the handling for NumberAdd/Subtract/Multiply as
well, which triggers surprisingly often in our core benchmark suites..

Bug: v8:8015
Change-Id: If8ec1bc82d1e1b71285c829262a0d343a4eb2af7
Reviewed-on: https://chromium-review.googlesource.com/1226033
Commit-Queue: Benedikt Meurer <bmeurer@chromium.org>
Reviewed-by: Jaroslav Sevcik <jarin@chromium.org>
Cr-Commit-Position: refs/heads/master@{#55943}