[FIX] ASR: use a real PSD square root in asr_calibrate

[sappelhoff]

asr_calibrate computed M = linalg.sqrtm(np.real(Uavg)), which only strips a pre-existing imaginary part but does not make Uavg positive-semidefinite. If the robust average covariance Uavg has even a tiny negative eigenvalue (realistic for average-referenced / rank-deficient EEG covariance means), scipy.linalg.sqrtm of the resulting indefinite symmetric matrix returns a complex M, which then silently propagates through eigh(M) into V and the threshold matrix T. asr_process already guards against this by taking np.real() of its eigen-quantities, but asr_calibrate did not.

Add _psd_sqrtm(), a real symmetric PSD square root that symmetrizes its input and clamps numerically-negative eigenvalues to zero before taking the root, and use it in asr_calibrate. For a genuinely PSD Uavg (the normal case) this is the same principal square root sqrtm would return, so calibration is numerically unchanged.

Testing

Added a regression test showing _psd_sqrtm stays real on a slightly indefinite matrix where sqrtm returns complex. Full tests/test_asr.py passes; ruff clean.

[codecov]

Codecov Report

✅ All modified and coverable lines are covered by tests.
✅ Project coverage is 83.25%. Comparing base (f71758d) to head (3183ac1).
⚠️ Report is 1 commits behind head on master.

Additional details and impacted files

@@            Coverage Diff             @@
##           master     #114      +/-   ##
==========================================
+ Coverage   83.22%   83.25%   +0.03%     
==========================================
  Files          25       25              
  Lines        2843     2849       +6     
==========================================
+ Hits         2366     2372       +6     
  Misses        477      477              

☔ View full report in Codecov by Harness.
📢 Have feedback on the report? Share it here.

🚀 New features to boost your workflow:

• ❄️ Test Analytics: Detect flaky tests, report on failures, and find test suite problems.

[sappelhoff]

Note: includes the nonlinear_eigenspace reproducibility fix from #118. test_asr_class[*-riemann] compares incremental-vs-bulk output with a loose tolerance and is flaky on CI because that path used an unseeded RNG; this can trip on any PR regardless of what it changes. #118 is the root fix — once it merges first, this commit drops out on rebase.