ARROW-13937: [C++][Compute] Add explicit output values to sign function and fix unary type checks

[edponce]

This PR adds explicit output values to the sign function for the unsigned integer kernel. This makes the kernel more readable and consistent with the other sign kernels.

Also, enable_if checks are fixed for scalar unary arithmetic functions (sign, negate, absolute, ceil/floor/trunc).

[github-actions]

https://issues.apache.org/jira/browse/ARROW-13937

[github-actions]

⚠️ Ticket has no components in JIRA, make sure you assign one.

[edponce]

After revisiting the sign compute function, I noticed that kernels are registered with Int8Type for integral types and corresponding floating-point for floating-point types. The kernels' Call methods are stated to be enabled if they return unsigned int, signed int, and floating-point. The unsigned int kernel is therefore never invoked. To fix this, the enable_if checks should be done against Arg0 or Arg1.

Also, shouldn't the floating-point kernels also return the same int type for sign function?

[edponce]

@lidavidm Please help with reviews when you get a chance.

[lidavidm]

Also, shouldn't the floating-point kernels also return the same int type for sign function?

np.sign returns the same type as the input.

[lidavidm]

Thanks. Having test coverage info would help us realize that overloads/branches aren't being taken, I think.

[edponce]

cc @lidavidm The floating-point kernels need to return a floating-point type because of sign(NaN) = NaN and NaN is not support in integer types. For the integral kernels, Int8 is returned because output can be any of (-1,0,1).

I was thinking last week of tests for code coverage. I actually know of several snippets of dead code.