[FEA] Support different index/stride types

[cliffburdick]

index_t is currently used for both index and stride computations, and users select which one they want at compile time. 64b stride/index is the default. In most cases users would never need a 64b index, and it's wasted instructions using them. Most CUDA libraries don't even support 64b indexes and only strides anyways, so it doesn't make sense to allow this.

This feature is to create a second type offset_t that is used for stride calculations only, and the user can select which they want to use.