Add ShareClaw logistic regression label projection variants

[anubhav1004]

Summary

• add shareclaw_logreg_scaled, a scaled PCA logistic regression baseline
• add shareclaw_logreg_balanced, a class-balanced scaled PCA logistic regression baseline
• keep both methods lightweight, reproducible, and task-compatible

Validation

• viash test src/methods/shareclaw_logreg_scaled/config.vsh.yaml
• viash test src/methods/shareclaw_logreg_balanced/config.vsh.yaml

Small-resource notes
On the official resources_test datasets used locally on the VM:

• shareclaw_logreg_scaled improved accuracy and f1_weighted over stock logistic_regression on cxg_immune_cell_atlas
• shareclaw_logreg_balanced improved f1_macro over stock logistic_regression on cxg_immune_cell_atlas
• all logistic regression variants achieved perfect scores on the small pancreas test resource

These methods are intentionally simple additions meant to provide useful lightweight baselines with different metric tradeoffs.

[anubhav1004]

Added two stronger LDA variants to this branch after a ShareClaw-guided sweep over the official small task resources.

Current best small-resource candidate:

• shareclaw_lda_svd: aggregate_macro_f1=0.700112 across cxg_immune_cell_atlas + pancreas
• On cxg_immune_cell_atlas: accuracy=0.546012, f1_macro=0.400224, f1_weighted=0.562549

Current best Zebrafish local baseline:

• shareclaw_lda_lsqr_auto: macro_f1=0.540053, accuracy=0.394122, balanced_accuracy=0.697012, weighted_f1=0.422995

Both new methods passed viash test on this branch.

[anubhav1004]

Hi @rcannood — would love your review on this when you get a chance! The two logistic regression variants show improved accuracy/F1 over the stock baseline on cxg_immune_cell_atlas. Both methods pass viash test. Happy to address any feedback. Thanks!