Add SGD constant fitting for symbolic regression

[alcides]

Summary

• Add an optional sgd dependency extra for JAX, Equinox, and sympy2jax.
• Add fit_constants to optimize numeric constants in a SymPy expression with full-batch gradient descent.
• Include a runnable example and tests for coefficient recovery and argument validation.

Validation

• uvx ruff check . --fix
• uv run --extra sgd pytest tests/geml/constant_optimization_test.py -q (4 passed)
• uv run pytest -q --ignore=tests/gp/performance_test.py (503 passed, 8 skipped)
• Full uv run pytest -q was stopped during existing performance benchmarks after reaching 81%; unrelated benchmark tests were taking several minutes.

Closes #52