- Remove dtype state from reduction.
- Introduce preamble_generators.
- Introduce function_manglers, to replace result_dtype_getters
- Introduce defines.

- Rename apply_subst to expand_subst.

- Track if precomputes in other substitution rules are uniformly
  in-footprint, and if so, perform substitution to storage reference
  even within substitution rules.

- Remove 'formal argument' sweep spec from precompute(). This is purely
  counterintuitive, because it finds an intermediate set of inames that
  then get swept out and thereby may pick up much more than the formal
  argument referenced.

- Fix the 'in-footprint' criterion to be the storage-to-sweep map with
  the sweep dimensions projected out.

insn.all_inames() -> kernel.insn_inames(insn)
Scheduling:
- Only schedule referenced inames.
- Only work towards insns that have their dependencies satisfied.