Implement cuda::(std::)memcpy

[jrhemstad]

I would like an optimized cuda::memcpy function that takes a CG to do the copy and handles alignment and doing vectorized copies when possible.

Much of the machinery is already available in cuda::memcpy_async, but it needs to be exposed as a public cuda::memcpy function.

[brycelelbach]

As mentioned, a lot of the machinery exists. We still need to do peeling, I think, and validate codegen.