Reduces allocation for the test case I was looking at by about 1.2%.
Mostly from avoiding allocation of some folding functions which turn
into let-no-escape bindings which just reuse their environment instead.

We also force inlining in a few key places in CmmSink which helps a bit
more.