Fix regression in action graph games

[tturocy]

This fixes a regression in action graph games, that was causing a segmentation fault due to not initialising the strategy index correctly after doing so was introduced in a refactor a few months ago.

I have tested this with 2x2.agg and the previous behaviour has been restored.

While we are not planning to do anything major with action graph games for 17.0, this would be a good opportunity for us to add some entries into the test suite just to guard against regressions on the limited set of AGG files we currently have.

[rahulsavani]

I can confirm that the fix seems to work fine.

I will now add tests and then ask @tturocy to review those.....

[tturocy]

I've now added support for outcomes in (Bayesian) AGGs in Python, which indirectly solves the .to_arrays() problem (and also #871)

[tturocy]

@rahulsavani In view of the discussion in #879, would it make sense only to run tests in floating-point for the purposes of considering the regression identified in this PR as "fixed"? The broader issue of the semantics of numerical data types in (B)AGGs has always been there (we have just never thought carefully about it), so the fix in this PR would at least restore us to the world before the bug was introduced.

The plan in #879 would then point towards further tests to add in due course.

[rahulsavani]

@rahulsavani In view of the discussion in #879, would it make sense only to run tests in floating-point for the purposes of considering the regression identified in this PR as "fixed"? The broader issue of the semantics of numerical data types in (B)AGGs has always been there (we have just never thought carefully about it), so the fix in this PR would at least restore us to the world before the bug was introduced.

The plan in #879 would then point towards further tests to add in due course.

Absolutely, that's what I had in mind and planned to discuss with you. I will update the tests accordingly asap.