lattice: Restore `tmerge_fast_path` for optimizer lattice (#50291)
This function is slightly misnamed, because, while it does perform
some fast pathing, the rest of the tmerge functions assume it has
been run. When I removed the OptimizerLattice in #50257, we
accidentally lost the call to `tmerge_fast_path` in the optimizer_lattice
path, regressing tmerge quality in the optimizer (which admittedly
we don't use a lot, but I did just add another use for in #50287).
The proper fix is probably to split the tmerge into a helper that
always calls the fast path and then delegates to the current
implementation for everything else, but for now, just try moving
the fast path to the start of the PartialsLattice, which should
restore correctness and have reasonable performance, since
there's only a handful of pointer checks between the entry to the
inference tmerge and the partials lattice.