Add teacher-student policy distillation colab for AlphaZero on Tic-Tac-Toe

[bbrianxiao]

Add a teacher-student policy distillation colab (AlphaZero, Tic-Tac-Toe)

Adds open_spiel/colabs/alpha_zero_policy_distillation.ipynb: trains an AlphaZero
teacher from scratch, distills it into a much smaller network, and measures how
much strength survives. Uses the existing alpha_zero.alpha_zero() loop and the
existing model_nnx.py Model for both teacher and student.

agent	params	vs minimax	exact regret
random baseline	-	0.055	0.4212
student (distilled)	3,338	0.485	0.0872
teacher-raw	338,698	0.485	0.0867
teacher+MCTS	338,698	0.500	0.0033

• Student matches its teacher at 101.5x fewer parameters (0.0872 vs 0.0867).
• Search accounts for a ~26x blunder reduction that win/draw/loss cannot see —
  those two agents differ by 3 games in 100, and the direction flips between runs.

[alexunderch]

Good one! Clear ipython outputs if possible, please