Add matrix, array and matrix multiplication for integer64

[hcirellu]

This pull request adds support for matrix, array and %*% for integer64. Since there are no S3 methods for matrix and array, they are added.
In matrix and array the base functions are used. In order to display convinient warnings and errors that show the call (e.g. matrix(1:10, nrow=5) instead of base::matrix(...)) I use withCallingHandlers(). If there is a generic call, this is displayed, and if not, the S3-method call is displayed.
In order to display matrices and arrays consistently to R, str.integer64 is changed accordingly.
colSums.integer64 and rowSums.integer64 are refactored to throw R consistent error messages.
as.matrix.integer64 is added to coerce accordingly.
Since aperm.integer64 isn't exported, one has to use the generic. Thus it is refactored to use NextMethod().
The matrix multiplication uses a new helper function target_class_for_Ops(), which determines the class of the result. If a matrix multiplication of integer64 and complex is performed, I expect the result in complex. It also throws R consistent error messages for Ops. This helper is also intended to be used for other Ops methods like *.integer64 in the future. The remaining method checks the matrix dimensions and performs the calculation.

Closes #45

[MichaelChirico]

I think it's basically ready. Thanks!

[hcirellu]

Thank you for the review!

[hcirellu]

I added this as.integer64, because I think it is good practice to assure that this method returns integer64, even though it might have been called directly internally without integer64.
But here the result would be inconsistent within the method, because if the input would be a non-integer64 matrix, it would be returned without converting to integer64.
Should I modify it to assure an integer64 return value or leave it as it is?
@MichaelChirico ?

[MichaelChirico]

Generally I think it's overkill to call as.* again from within a .* S3 method -- if we get non-integer64 input, user is playing with fire :)

We have a few ugly checks in as.data.table.data.frame that are basically there for back-compatibility -- I'd rather the user be stuck debugging such cases:

https://github.com/Rdatatable/data.table/blob/a325db9e0c6a0cd1076c086d5fc17a548bccd1a5/R/as.data.table.R#L254-L261