v2.3.0

fix(bench): make CodSpeed suite Valgrind-safe

Two benchmarks have never completed under CodSpeed's Valgrind
instruction-counting mode:

* test_poly_jit_eval_deg10 — failed because numpy was not installed
  in the CI venv, and even with numpy the 1 000 000-point evaluation
  would be prohibitively slow under Valgrind.  Replace with an inline
  variant that evaluates at 1 000 points and skips gracefully if numpy
  is absent.  Add numpy to the CI pip install.

* test_sparse_interp_multivar_10vars — consistently timed out the
  2-hour job (10 variables × heavy oracle calls under Valgrind).
  Renamed to test_sparse_interp_multivar_2vars; uses size=2 (2 vars,
  3 terms) which completes in milliseconds while still exercising the
  multivariate path.  Large-size runs remain available via
  cas_comparison.py for wall-clock comparisons.

Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>