Return full profiles from enumpoly_solve

[tturocy]

This proposes a change to enumpoly_solve such that it now returns only full profiles.

By convention, it returns profiles which specify only pure actions on information sets which are reached with zero probability.

[rahulsavani]

@tturocy, an update:

• the two tests you marked as xfailing in Restrict search space for enumpoly to be strictly interior. #764 have been updated and I don't see any issue with those;
• the earlier two xfailing examples based on the trust game have naturally been fixed as expected;
• I simply commented out the only other two xfailing tests, which are both Stephen's large (huge) payoff game, and don't really make sense in the double setting;
• I added 3 new test cases for the "two off the path" aspect of things (see test_enumpoly_behavior_{10,11,12}) -- with these I am happy that what is being done is reasonable enough, but I am not sure we discussed this all explicitly, so see if you are happy where the centroid is being used (for test_enumpoly_behavior_10) I marked the relevant supports from the CLI in the comments for your convenience.

[tturocy]

Seems like this is looking pretty good. Things to check before we merge and draw a line under this:

• Is this now consistent with what lp and lcp do - centroid for two-step deviations? Acknowledged that there is going to be a difference in one-step deviations, because lp and lcp might in principle find the "mixed" end of a continuum of equilibria

• Are we now documenting this well, and specifically that we follow this centroid convention consistently across the three sequence-form methods?

[rahulsavani]

Seems like this is looking pretty good. Things to check before we merge and draw a line under this:

• Is this now consistent with what lp and lcp do - centroid for two-step deviations? Acknowledged that there is going to be a difference in one-step deviations, because lp and lcp might in principle find the "mixed" end of a continuum of equilibria
• Are we now documenting this well, and specifically that we follow this centroid convention consistently across the three sequence-form methods?

For the first bullet: I believe so yes. In relation to this I added a 2-player test case with a 2-deviation infoset for {lp,lcp,enumpoly}_solve.

For the second bullet, no, we need to improve the docs. Let's discuss how to explain this for enumpoly, as that needs a little care.