Integer data type

[timmoon10]

As of d2c414a1846b072ea6607c85a0254c5b1277d4e7, int is the standard integer data type. However, we may need 64-bit integers for very large matrices. Variables susceptible to overflow should be changed to El::Int. As a rule of thumb, indices into Elemental matrices should be El::Ints.

[ndryden]

With these changes, I'm getting a lot of warnings about narrowing casts.

[ndryden]

(Background/refresher: fundamental C++ types and fixed-width integer types)

Just to add some context for this discussion, these are our main sources/issues with integers, as far as I can recall:

• Elemental defines El::Int to be either int or long long depending on our compile-time flags. (There is also El::Unsigned.) We want to use El::Int when interacting with Elemental matrices.
• The C++ STL (e.g. std::vector) tends to use size_t for just about any size-related quantity. This can lead to us either narrowing a size_t or comparing signed/unsigned integers.
• cuDNN expects parameters to be int.
• MPI expects parameters to be int.

Ideally, we should come up with a consistent use of integers that satisfies all of these.

Edit: An additional thought: while we don't want to do it for production, we could compile with -ftrapv, which will trap for signed overflow on addition, subtraction, and multiplication.