Change warmup strategy for sliding window

[dsharletg]

This branch changes how we handle the "warmup" of sliding windows before running the "steady state". The old strategy was to make the first iteration handle the warmup. This branch changes the strategy to "back up" the loop so the steady state warms up the loop itself, inserting the necessary ifs to guard the producer and consumer to avoid producing or consuming out of bounds values.

This change has the following impacts:

• It is now possible to generate much finer grained sliding window schedules with low overhead. The 'blur' app in this branch is ~30% faster than master.
• Some sliding window schedules appear to get possibly a bit faster in some test cases (local_laplacian, stencil_chain, camera_pipe).
• Folded storage buffers can be smaller, because the folded storage only needs to accommodate the steady state of the loop. The warmups in the old strategy needed larger folded storage windows. See the changes to camera_pipe and local_laplacian for example. This should make storage folding less confusing (the reason for the larger factors was confusing and seemed unnecessary).

This avoids one of the motivations for #5204, and is hopefully helpful for #1820

[dsharletg]

Reopening now that those two prerequisite PRs have been merged.

[steven-johnson]

Looks clean. Ready to land?

[dsharletg]

I think this is ready to merge, @abadams, @rootjalex did you want to take a final look at this before merging?

[rootjalex]

The interval code lgtm