FIX Handle zero-norm blocks in proximal operators

[ktz03]

Context of the PR

Zero-norm blocks cause the radial proximal operators to divide by zero. L2_05 raises ZeroDivisionError, while BlockMCPenalty and BlockSCAD produce NaNs that can propagate through a multi-task fit. Block soft-thresholding also fails when both the block norm and threshold are zero, including unregularized L2_1 and GroupLasso fits.

For example, an orthogonal design with a nonzero response on one row and zero responses on the other rows reaches this case:

import numpy as np
from skglm import GeneralizedLinearEstimator
from skglm.datafits import QuadraticMultiTask
from skglm.penalties import BlockMCPenalty
from skglm.solvers import MultiTaskBCD

X = np.asfortranarray(np.sqrt(3) * np.eye(3))
Y = np.asfortranarray([[3., 4.], [0., 0.], [0., 0.]])
estimator = GeneralizedLinearEstimator(
    datafit=QuadraticMultiTask(),
    penalty=BlockMCPenalty(alpha=0.1, gamma=4.),
    solver=MultiTaskBCD(fit_intercept=False, use_acc=False,
                        max_iter=5, max_epochs=20),
).fit(X, Y)
assert np.isfinite(estimator.coef_).all()

Contributions of the PR

Return zero for zero-norm blocks before radial scaling, and include the zero-threshold boundary in block soft-thresholding. Add regression tests for plain and compiled penalties and dense/sparse multi-task fits, comparing against the same problem without the zero-response features.

Validation on Windows with Python 3.12.14, NumPy 2.2.6, SciPy 1.15.3, Numba 0.61.2 and scikit-learn 1.7.2:

• The new tests give 22 failures and 2 passing controls on the unchanged implementation; all 24 pass after the fix. The full penalty module passes 43 tests.
• The complete pytest -q skglm/ run passes 357 tests with 17 existing xfails; the pristine tree passes 333 with the same 17 xfails.
• 304 direct proximal-operator checks against a radial objective oracle pass, including nonzero controls and zero/positive regularization. Eight dense/sparse, positive/unconstrained GroupLasso checks pass against the orthogonal-design solution.
• Whole-codebase flake8 and pydocstyle commands from the lint workflow pass.
• The documented html-noplot build exits successfully on both trees with the same existing diagnostics (Sphinx reports 56 warnings, including the existing index.rst:123 ERROR: Unexpected indentation). Example plots were not executed.

Other platforms, Python versions and performance were not tested. The existing float32 block soft-thresholding compilation issue is outside this fix.

Checks before merging PR

• added documentation for any new feature (not applicable: bug fix)
• added unit tests
• edited the what's new (if applicable)