Improve constant factors of injection into sums.

Right now we pay a performance penalty when using multiple effects in parallel, even though they fuse correctly, because `inj` (which is called every `send` is O(n) in the number of effects rather than O(1)). A smarter `Sum`/`Union` type could perhaps ameliorate this situation, though we may have to change some code upstream.

Start by tracing `inj`, which is called by every `send`, and inspect the `Sum`/`Union` types and any upstream code they depend on. Done means reducing injection from O(n) to O(1) for multiple parallel effects without breaking their fusion behavior, with performance evidence and relevant tests.

Written by the indexing model from the issue text.