ARM: Do not set FPSCR when converting to clamped uint8

ARM: Do not set FPSCR when converting to clamped uint8

Setting the FPSCR flags is expensive on some CPUs. Get rid of repeated
setting of the FPSCR by relying on the correct default flags being set
when doing uint8 clamping. Also use vcvt_u32_f64 instead of vcvt_s32_f64,
which enables removing the check against zero (vcvt_u32_f64 will clamp to
zero).

To be on the safe side, add asserts to check that the VFP rounding mode
flags are set to default as expected.

This increases performance of a hot loop repeatedly setting
Uint8ClampedArray values on some CPUs by as much as a factor of 12.

BUG=v8:3253
LOG=N
R=jacob.bramley@arm.com, rmcilroy@chromium.org, ulan@chromium.org

Committed: https://code.google.com/p/v8/source/detail?r=20676

Review URL: https://codereview.chromium.org/230473005

Patch from Olli Etuaho <oetuaho@nvidia.com>.

git-svn-id: http://v8.googlecode.com/svn/branches/bleeding_edge@20755 ce2b1a6d-e550-0410-aec6-3dcde31c8c00