CooperateISO

[adrianhutter]

Adding 3 strategies (LongtermTfT, ISO, CooperateISO) from https://arxiv.org/abs/2303.03519.

Made a few changes relative to the versions used in the paper. Most importantly, ISO now uses scipy.optimize instead of torch's Adam. This works equally well, is faster, and avoids having to import torch.

Ran the CooperateISO implementation from this PR against all strategies in Axelrod-4.14.0, and verified that it outperforms EvolvedLookerUp2_2_2 and DBS at both noise levels 0% and 10%.

Fix Read the Docs build failures by updating documentation requirements and conf.py:

• Pinned sphinx to <9.0.0 and added sphinx-rtd-theme in docs/requirements.txt to ensure compatibility and prevent theme lookup errors.
• Removed scipy from the mock modules list in docs/conf.py so that submodules like scipy.optimize can be correctly imported from the installed package during documentation generation.