perf: cache lambdified equations and avoid redundant export work

[M-Colley]

Summary

CallableEquation._lambda was a @property that called sympy.lambdify on every access, so every predict() and every direct equations_["lambda_format"](X) call re-generated and exec'd Python source before evaluating the expression. This caches it with functools.cached_property; __getstate__ drops the cached function, which cannot be pickled, and it is rebuilt lazily after loading.

Smaller redundancies on the same paths:

• latex() called refresh() and then sympy(), which refreshes again, so it rebuilt every export twice.
• idx_model_selection filtered with DataFrame.query(f"loss <= {threshold}") on every predict(). A boolean mask is about 3x cheaper and avoids round-tripping the float through a string.
• add_export_formats and _search_output_to_callable_expressions deep-copied frames they only read, and the former rebuilt the SymPy symbol list for every row.
• _run made a full copy of X just to inspect its dtype, and another before jl_numpy_array, which already copies into Julia memory. It now uses np.asarray, so a copy is only made when the dtype changes.

Measurements

These use a model loaded with from_file from a 6-equation hall of fame with 5 features, on Windows with Julia 1.11 and SymbolicRegression 2.7.0. The machine was shared, so absolute times are noisy.

call	master	this PR
equations_.iloc[-1]["lambda_format"](X) (10 rows)	3.44 ms	0.15 ms
predict(X) (10 rows)	8.8 ms	5.5 ms
latex()	25 ms	17 ms

Calling lambdify alone on a 5-variable expression takes about 1.75 ms, against about 0.008 ms to evaluate the result.

Tests

• test_lambda_format_compiles_once_and_stays_picklable: lambdify runs once across repeated calls and predict(), and the callable still round-trips through pickle. On master lambdify runs 6 times.
• test_latex_refreshes_once: on master refresh runs twice.

Both new tests fail on master and pass here. TestBest, TestLaTeXTable and the export pickling tests pass locally.

This came out of a broader review of the codebase. The related PRs are independent of each other and merge cleanly together.

🤖 Generated with Claude Code

[MilesCranmer]

Thanks for this! I don't see a user-facing problem this PR fixes, and it adds code to maintain. Closing for now; happy to revisit if you hit a real slowdown.