Address some of the fusion issues in #202 - #241

it would previously replace `takeWhileEnd p` with

    S.reverse . S.takeWhile p . S.reverseStream

But `S.reverse` can only fuse with its argument, and `S.reverseStream`
only fuses with its consumer. So all possible fusion is within this
pipeline, and in the best case we obtain a direct implementation.

There might be some value in this if we had a
`S.reverseStream/S.reverse` rule, so that in multiple chained invocations
of `takeWhileEnd` we could eliminate some intermediat steps. But no such
rule exists at the moment.

There is a comment that

> length needs to be phased after the `compareN/length` rules otherwise
> it may inline before the rules have an opportunity to fire.

but these rules are marked `[~1]`, so we can inline `length` in phase 1
to expose the `stream` in its definition and actually allow it to fuse.

This mirrors commits 1e10823, just for
strict functions.

as it does a case match.

because `dropWhileEnd` is not.

it was matching on `singleton`, but `singleton` gets rewritten to
`s.singleton`. An additional rule makes this robust.

because the rules for `dropWhile` are only enabled before phase 1