Add diagonal covariance option for dual estimation + hmm_features

[bengriffin3]

Summary

This PR adds support for diagonal covariance matrices in the HMM dual estimation pipeline, matching the existing capability in the main HMM model (config.diagonal_covariances). This enables session-specific re-estimation of variance-only observation models, useful for parcellations or analyses where cross-channel covariance structure is not required or not identifiable.

Key changes

• Extend hmm_dual_estimation (in analysis/post_hoc.py) with a new argument diagonal_covariances=False.
• Implement efficient diagonal-only covariance estimation (weighted per-channel variances).
• Ensure returned matrices remain full (n_channels × n_channels) with off-diagonals zeroed.
• Update hmm_features to:
  • pass through the new option,
  • extract diagonal terms when diagonal_covariances=True,
  • maintain consistency with optional partial correlation transforms.
• Update Model.dual_estimation to respect config.diagonal_covariances.
• Update and expand relevant docstrings.

Motivation

Enable dual-estimation workflows where only state-specific variances are needed (e.g., amplitude/power-driven dynamics), consistent with the diagonal-covariance option already available during HMM training. This brings feature parity to post-hoc analysis functions.

Notes

• Full-covariance behaviour is unchanged.
• The diagonal option is opt-in and affects only explicit uses.
• Tested on synthetic data to ensure off-diagonal entries remain zero and behaviour is correct.

[cgohil8]

Nice! Looks good. Maybe one minor comment, can you change the order of the arguments to:

def hmm_dual_estimation(
    data,
    alpha,
    zero_mean=False,
    diagonal_covariances=False,
    eps=1e-5,
    n_jobs=1,
):