Streaming prefetch in add_prefetch/precompute

Adds a "streaming" option to add_prefetch and precompute, which modifies precompute to copy any overlapping local data between iterations of the outer loop of a split iname, i.e. to avoid unnecessary redundant global reads. At the moment this manifests as an optional stream_iname parameter (specifying the outer iname) to add_prefetch and precompute which, when not None, implements the streaming version. (There are a number of spots in my implementation that possibly could/should be improved.)

⚠ Depends on !284