optimize float<->fp16 conversion on aarch64

https://gcc.gnu.org/onlinedocs/gcc/Half-Precision.html says there are hardware instructions to do this conversion, but only with certain compiler options. I'm not seeing those options in the build log on CI.

This came up because enabling this platform in CI had the effect of blocking 5.10 integrations for a while. Hopefully doing an explicit cast will be enough: https://codereview.qt-project.org/#/c/211734/ but it probably still would be a good idea to ensure these conversions are done as efficiently as possible.

The older patch which added initial support is https://codereview.qt-project.org/#/c/185724/

A build log where I saw the warning (which is an error because of -Werror): https://testresults.qt.io/logs/qt/qtbase/ae93e48e3f326b89adda3ee171d2aff437f9d869/LinuxUbuntu_16_04x86_64LinuxQEMUarm64GCCqtci-linux-Ubuntu-16.04-x86_64-1-e401faDeveloperBuild/2a48a3b7dcb9652b1532a3ad14de7580b5b9b1ee/build_1510903300/log.txt.gz