ARROW-1562: [C++] Numeric kernel implementations for add

[fexolm]

It's just an initial AddKernel implementation (work in progress). I could not come up with a better implementation of type inference, so I want to hear your thoughts about implementation.

[fsaintjacques]

Multiple notes:

1. The Compare kernels only support comparing the same type, it expects the caller to decide (via Cast kernel) how to deal with mis-matching types. I think this would be the most simple and sane behavior.
2. The Sum kernels upcast to the widest type. This might be less of an issue here, but I think we should take a decision.
3. I wouldn't merge this without supporting null, this is too frequent to avoid.
4. This code is regular enough to support other basic arithmetic operations, i.e Addition, Substraction, Multiplication, Division.
5. The code is also regular enough to support (array, scalar) operation. See how this is implemented for Compare kernels.

[emkornfield]

@fexolm it looks like @fsaintjacques gave some feedback on implementation do you need more guidance?

[fexolm]

@emkornfield no, I just can not find the time to finish this. I will return to this as soon as I get time.

[fexolm]

@fsaintjacques, @emkornfield
I've removed type inference from the code, also add nulls support.
BTW I haven't decided yet how it can be architecturally generalized to support other arithmetic operations. I see, that almost all the code could be reused in other operations, but not sure how to implement it.
Maybe it would be better to do this in other pr, because I'm not sure I have enough time for it now.

[emkornfield]

Thanks @fexolm@fsaintjacques do you have time to review?

[fsaintjacques]

Will review again.

[fsaintjacques]

I created https://issues.apache.org/jira/browse/ARROW-7017 as a followup.