[ParitKansal]

Closes #59.

Adds symbolic (partition-based) estimators following Cafaro, Lord, Sun & Bollt, "Causation entropy from symbolic representations of dynamical systems", Chaos 25, 043106 (2015), arXiv:1507.07262.

What is added

• symbolize (entropy.py): partitions each variable into n_symbols cells (Eq. 2–3 of the paper). Options: quantile (equal-frequency; edges placed between distinct values so tied values always share a cell), uniform (equal-width), precomputed (data already symbolic).
• symbolic_entropy, symbolic_mutual_information, symbolic_conditional_mutual_information: plug-in estimators (Eq. 13–15), in nats like the other estimators.
• method="symbolic" in conditional_mutual_information (new explicit args n_symbols, partition; no **kwargs).
• information="symbolic" in discover_network (uses the default 2-symbol median split).
• README and CHANGELOG entries.

Tests (24 new, tests/core/information/test_symbolic.py)

• CMI matches Eq. (15) computed term by term; chain X→Z→Y gives I(X;Y|Z)=0.
• Logistic map with the generating partition at 1/2 gives memoryless fair-coin symbols (the setting of Sec. IV of the paper).
• Tied/count data (e.g. 70% zeros) keep both symbols.
• discover_network(information="symbolic") recovers exactly the true edge X0→X1 at lag 1.
• Full suite: 497 passed; black, isort and flake8 (CI settings) clean; 100% line coverage of the new functions.

Cross-checks against other libraries (random cases, differences at rounding level)

Quantity	Reference	Max abs. difference
Entropy	pyitlib, infomeasure	4e-16
MI	sklearn mutual_info_score, pyitlib	2e-15
CMI (incl. 2-column Z)	pyitlib, infomeasure	4e-15
CMI	tigramite CMIsymb	3e-15
Bins	sklearn KBinsDiscretizer	uniform identical; quantile equal-frequency (only exact-edge points differ)

Open questions for maintainers

1. Should discover_network expose n_symbols/partition? That would mean threading two explicit parameters through the internal functions, so I left the default (2 symbols) for now.
2. No small-sample bias correction (e.g. Miller–Madow); the permutation test uses the same estimator, so the bias partly cancels. Could be a follow-up.