NumCosmoXCOR_KERNEL_MAX_ELL_BLOCK

Hard cap on the number of multipoles in a single get_eval_vectorized() call: NcXcorKernel‘s internal per-block state uses fixed-size stack arrays sized by this constant (not just the Levin integrator’s own, larger ell_cache_max). Exceeding it is a fatal, non-catchable g_error. Public so callers planning $\ell$-block tilings (e.g. NcXcorSolver) can respect it without duplicating the number.