Performance: pseudo_inverse_stacked has significant bottlenecks with large matrices

[wilczman]

Problem

The pseudo_inverse_stacked() function in cedalion has two major performance bottlenecks that cause significant slowdowns when working with large sensitivity matrices:

1. Duplicate matrix multiplication: Lines 37-38 compute AAt = Adot.values @ Adot.values.T twice
2. Inefficient eigenvalue calculation: Uses full eigendecomposition np.linalg.eig() when only the largest eigenvalue is needed

Current Code Location

src/cedalion/imagereco/solver.py -> pseudo_inverse_stacked() around lines 37-38 and eigenvalue computation

Proposed Solution

1. Remove the duplicate AAt computation
2. Replace np.linalg.eig(AAt)[0][0].real with np.linalg.norm(AAt, ord=2) for largest eigenvalue

[wilczman]

Fix in PR: #123