Add Linear Regression Rereferencing (LRR)

[cboulay]

Summary

• Add a self-supervised regression framework (SelfSupervisedRegressionTransformer) and a concrete LRR implementation (LRRTransformer) in ezmsg/learn/process/ssr.py
• LRR predicts each channel from the others in its cluster via ridge regression and subtracts the prediction: y = X @ (I - W)
• The effective weight matrix is delegated to AffineTransformTransformer, which automatically exploits block-diagonal structure when channel_clusters are provided
• Bump ezmsg-sigproc dependency to >=2.13.1 to use AffineTransformTransformer.set_weights

Details

Framework (SelfSupervisedRegressionTransformer):

• Accumulates channel covariance C = X^T X and solves per-cluster ridge regressions via the block-inverse identity (one matrix inverse per cluster instead of a per-channel Cholesky loop)
• Supports incremental and batch fitting modes
• All linear algebra stays in the source array namespace (NumPy, CuPy, etc.)
• Subclasses implement _on_weights_updated and _process

LRR (LRRTransformer / LRRUnit):

• _on_weights_updated computes I - W and passes it to an internal AffineTransformTransformer
• Subsequent fits use set_weights for fast in-place updates without a full state reset
• Supports pre-calculated weights from a numpy array or CSV file path
• LRRUnit provides the ezmsg Unit wrapper with INPUT_SAMPLE subscriber