Make sure default compile flags are used correctly and are well documented

[denisalevi]

• We are using -use-fast-math, which uses intrinsic math operations at the expense of floating-point precision. Needs to be documented and needs a prefs to turn this off if someone need the precision. Maybe also quantify the precision loss?
• Check all the optimization flags. This stackoverflow post mentions that -O0 will be passed only to the host compiler and that one needs to pass -Xptxas -O0 to turn of device code optimization. Double check that preferences are clearly documented and flags are passed to the correct compiler.