Fix NEON min/max for nan values.

The NEON intrinsics vmin/vmax always return nan when comparing to numbers rather than the first argument, as specified in the standard for std::min/std::max. To get around this, we use a combination of greater-than/less-than (which do adhere to the standard for comparing values against nan) and select.

Tested:

aarch64-linux-gnu-g++ -static -I./ -Iunsupported/ -Itest/  unsupported/test/cxx11_tensor_expr.cpp -o cxx11_tensor_expr
adb push cxx11_tensor_expr /data/local/tmp/
adb shell /data/local/tmp/cxx11_tensor_expr

Fixes #1937 (closed)