Bug in iNKA covariance

[sachaguer]

Related to Issue #290 , this PR fixes a bug in the iNKA covariance. The analytical based noise bias computation was missing a decoupling step leading to a misestimated noise bias and likely larger error bars. The updated functions is running for cross-validation with Gaussian simulations from #286 .

Ready to be merge as soon as the cross-validation is performed. In the meantime, design and runtime review is welcomed.

[sachaguer]

@cailmdaley, following #286, this PR is ready for a review.

[cailmdaley]

The fix is correct. I traced the decoupling end to end. The coupled analytic noise from Nicola et al. is now decoupled with the correct workspace. Coupled and decoupled quantities are used consistently downstream. Cross-bins correctly get zero noise bias. Approve.

One request, on a pre-existing issue this fix now touches: the signal and noise grids are misaligned by one multipole. get_fiducial_cl builds the signal on arange(1, lmax+1), so index i holds C(ℓ=i+1). The NaMaster output is 0-indexed, so index i holds N(ℓ=i). The sum fiducial_cl + noise_bias_cl therefore adds N(ℓ) to C(ℓ+1). The effect is small while the noise is flat, but please unify the two grids on the NaMaster convention (arange(0, lmax), ℓ=0,1 zeroed) so the fix rests on aligned inputs.

— Claude on behalf of Cail

[sachaguer]

The best I found to fix this is to add a zero monopole and drop the last value of the fiducial.
CAMB crashes when computing a theory at ell=0 so it seems to be the best solution to me. Happy to do something else if someone has a better idea.