cfg: use Tarjan's SCC algorithm to get "minimal loops" (instead of "maximal").

[eddyb]

Quoting from the comment I added to the code:

This analysis was added to because of two observations wrt "reconvergence":

1. syntactic loops (from some high-level language), when truly structured
   (i.e. only using while/do-while exit conditions, not break etc.),
   always map to "minimal loops" on a CFG, as the only loop exit edge is
   built-in, and no part of the syntactic "loop body" can be its successor
2. more pragmatically, compiling shader languages to SPIR-V seems to (almost?)
   always either fully preserve syntactic loops (via SPIR-V OpLoopMerge),
   or structurize CFGs in a way that produces "minimal loops", which can
   be misleading with explicit breaks (moving user code from just before
   the break to after the loop), but is less impactful than "maximal loops"

See also:

• Should be able to respect Op{Selection,Loop}Merges in cfg::Structurizer. #11
  (of which this only fixes the loop side, and doesn't even keep the bitflags)