Optimize more kinds of copies

[dsharlet]

Currently, we understand and optimize a few kinds of copies by expressing the copy as a buffer metadata manipulation like setting the stride. This is supported by this helper:

slinky/builder/optimizations.cc

Line 34 in 81bd036

bool is_copy(var src, expr src_x, int src_d, var dst, var dst_x, int dst_d, expr& at, dim_expr& src_dim) {

To be more specific, the copies we support now are of the form:

• broadcasts: f(x) = g(C) where C is not a function of x.
• copies: f(x) = g(x + C)

We could support additional kinds of copies, like:

• downsamples: f(x) = g(x * C), by multiplying the stride by C. Interestingly, we can't support upsamples, because dividing the stride is not right.
• "modular broadcasting": f(x) = g(x % C), we can only do this if we know that g doesn't already have a fold factor, so we can set the fold factor to C.

[Aelphy]

#343 seems to be stalled until fold factor + loop split issue is fixed