feat: Support pre-computed attributions in run_modisco via method='completed'

[jsdearbo]

Overview

This PR adds support for passing pre-computed attributions to run_modisco by introducing method='completed'.

Currently, grelu.interpret.modisco.run_modisco requires computing attributions dynamically from the model during the function call. While convenient for smaller workflows, this tightly couples the heavy sequence-to-function model inference with the motif discovery step.

Motivation

For larger analysis pipelines, it is often necessary to decouple attribution generation from MoDISco execution. For example:

• Compute constraints: Computing gradients for large models often requires heavy GPU allocation, while TF-MoDISco is typically run downstream on high-memory CPU nodes.
• Iterative downstream analysis: Researchers often need to run MoDISco multiple times on different subsets of the data using different masking windows or flank sizes. Re-computing attributions for every subset via the existing run_modisco workflow is computationally redundant and slow.

By allowing method='completed', users can load an existing numpy array of attributions and pass them directly into the MoDISco pipeline via kwargs.

Changes Made

• Updated src/grelu/interpret/modisco.py to bypass the internal attribution calculation when method='completed' is specified.
• Added standard ValueError handling if the attributions kwarg is missing.
• Added my name to AUTHORS.rst.
• Ran pre-commit to ensure compliance with Black/Flake8 styling.

Impact

This is a non-breaking change. The default behavior for other methods (inputxgradient, saliency, etc.) remains exactly the same.