Critical issues in cov utils

[nikosarcevic]

get_noise_power assumes one-digit bin indices (tr[:-1]), breaks for src10+ / lens10+

get_gaus_cov uses signal-only C_ell and adds noise^2 in one case; gaussian cov needs (C+N)(C+N) structure IIRC Current implementation misestimates cov

Casting ell to int16 risks overflow; use int32/float

Bandpower normalization uses np.gradient(ells) with unclear meaning--- likely inconsistent with delta ell / windowing.
get_SRD_cov hard-codes ncls=20 and assumes SRD ordering/dtypes; will silently mis-slice when bandpower count or ordering differs

TJPCovGaus.get_binning_info assumes integer unit-spaced ell edges and can crash if ell_edges missing so guard for that