(draft) Unary math cuda kernels

[jplauzie]

Part 2, along with #393. 3).

To complement the Time Quantity, here are some unary point wise math Quantity operators, that act on Quantities similar to Div, Mul, etc. This should allow a user to build up arbitrary functions f(x,y,z,t). To pick functions, I just went through the list of existing mumax3 math functions and grabbed ones that seemed likely to be used, and wrapped them into cuda kernels. All of the math functions I included had built in cuda functions except heaviside, so this is mostly just boilerplate to wrap them. The current list of functions is: sin, cos, tan, exp, log, abs, acos, acosh, asin, asinh, atan, atanh, cosh, sinh, tanh, erf, erfc, gamma, heaviside, pow, mul, atan2.

• To avoid colliding with existing math functions, I just stuck a Q on it for 'Quantity', but maybe there's a better naming scheme.

• I can always add/remove some functions, and there's also the possibility of adding any fused kernels if any seem particularly useful.

• I didn't include the PTX/wrapper_go files, apologies. If this gets merged, do you mind generating them? I'm still on a 10 series card, so I don't think I can generate it for the latest 50 series (Windows also seems to change the line endings from LR to CLRF compared to Linux)

• For some functions like pow, there are some choices about what to do if an input value is outside of the domain of the function. For now, I just followed what Div does, and it returns 0. This affects acos,acosh,asin,asinh,atanh,gamma. Tan might also need protection for poles. Pow also some custom logic so e.g. sqrt(-x) returns -sqrt(x).

• Stylistically: in cuda/unary_math.go I didn't quite follow the style of Div/Mul. Would it be better to revert this? I don't mind, it's just more verbose.

Included is a quick test file, that just picks a few random points and compares them to the existing math functions using Custom Quantities. It's kind of janky, saving a bunch of OVFs and loading from them, there is probably a nicer way to do this.