The Cvtqui2ss and Cvtqui2sd methods did overwrite the {src} register and
the given {tmp} register.
This CL fixes this to avoid the overwrite of the {src} register, and to
use the kScratchRegister instead of an explicitly passed tmp register.
It also adds u32 -> f32/f64 conversions.
All these methods can take either a register or an operand as src. This
sometimes saves a mov operation.

R=neis@chromium.org, ahaas@chromium.org

Change-Id: I912a2a3b9d9c356f04ce51869a01c6fd11c76dd3
Reviewed-on: https://chromium-review.googlesource.com/1017121Reviewed-by: Andreas Haas <ahaas@chromium.org>
Reviewed-by: Georg Neis <neis@chromium.org>
Commit-Queue: Clemens Hammacher <clemensh@chromium.org>
Cr-Commit-Position: refs/heads/master@{#52676}