Improved predicate splitter with MILP-based equivalence test

[robknight]

This branch started out as a replacement of the earlier "greedy" splitter with a MILP-based version, because the greedy version was unable to find feasible solutions to some of the longer custom predicates we've been testing with recently.

However, the MILP version also struggled with performance, sometimes taking 15 seconds or longer to return a solution.

What is implemented in this PR is a dynamic programming-based approach which is able to find feasible solutions in all cases I've tested with, with a MILP version implemented for testing purposes. There is a test which randomly generates sample predicates, and then verifies that the DP approach produces feasible solutions in all cases where MILP does. In addition, there are several new tests involving longer custom predicate chains.

The DP code is much faster than the MILP code for every scenario that I've tested. From the test:

Parity sweep: checked=1200, milp_feasible=1177, dp_divergences=0
  MILP total=73.158783697s mean=60.965653ms max=6.204155675s
  DP   total=4.329626493s mean=3.608022ms max=25.858044ms

So DP is ~16x faster at the mean, and ~240x faster at the maximum.

[robknight]

test

[ed255]

test