This MR adjusts SetLevels so that it is less eager to float a
HNF (lambda or constructor application) out of a lambda, unless
it gets to top level.

Data suggests that this change is a small net win:
 * nofib bytes-allocated falls by -0.09% (but a couple go up)
 * perf/should_compile bytes-allocated falls by -0.5%
 * perf/should_run bytes-allocated falls by -0.1%
See !12410 for more detail.

When fiddling elsewhere, I also found that this patch had a huge
positive effect on the (very delicate) test
  perf/should_run/T21839r
But that improvement doesn't show up in this MR by itself.

Metric Decrease:
    MultiLayerModulesRecomp
    T15703
    parsing001