'SequentialTag' was the wrong idea, it impeded sequential unroll.
Restrict-to-sequential and tagging have nothing to do with each other.
This is now realized in the code.

In particular, just touching a variable written to by a non-idempotent
instruction makes that instruction also not idempotent.

This allows easier access (during debugging) to the kernel that actually
gets scheduled.

This lets the higher-level routine choose its favorite, rather
than leave it out of options once the top choice doesn't work.

(not regular set, which isl stopped liking in 0.08)

The previous algorithm would miss hoist opportunities if a larger group
of candidate inames existed which then in turn did not lead to a viable
hoisted condition.