More efficient parallel search for simpdiv_solve

[tturocy]

The basic operation of simpdiv_solve is as follows:

• Fix a grid (triangulation) on the space of mixed strategy profiles
• Starting from an arbitrary point on that grid, find (roughly speaking) an approximate fixed point
• Refine the grid, and repeat.
• Do this until the resulting mixed strategy profile achieves a desired max_regret target.

Observe this method takes a mixed strategy profile as a starting point, and the equilibrium found (may, in general will) depend on the initial profile.

Currently a run of simpdiv_solve re-starts from an initial profile at each run. However, observe that this is not efficient! The core of the algorithm is a deterministic map (given a triangulation size) from an initial mixed strategy profile to a final mixed strategy profile. In practice, as we refine the grid size, different initial starting points wind up at the same final mixed strategy profile. At this point, because the algorithm is deterministic, the runs are the same for all subsequent grid size refinements.

Instead, it would make more sense to do this:

• Fix a grid on the space of mixed strategy profiles
• Take a set of initial points on that grid.
• For each point in the set, find the resulting point.
• Collect up the set of resulting points
• Refine the grid, and repeat.

In particular also note that the process at each grid size could be parallelised.

This would make a nice straightforward project, as it doesn't involve really digging into the simplicial subdivision code itself, but only the driving code. (The simplicial subdivision implementation itself is rather ugly and in need of a refactor, but that can be done completely independently of the idea above.)