docs: align Abacus wedge NPE workflow notes

[cursor]

Summary

• Updated existing Abacus/SBI workflow docs for current wedge NPE target/cache modes.
• Added concrete linear-increment cache and training examples, including _linear_eig.pkl, --increment_mode linear, --linear-increments, and --scale-invariant-features.
• Documented FlowJAX checkpoint/resume behavior and posterior class-probability diagnostics.
• Updated shared/eigenvalue_transformations.py module documentation to describe both softplus and explicit linear increment modes.
• Added a SCIENCE_LOG entry for the docs alignment.

Codepaths covered

• workflows/abacus_tweb/build_abacus_sbi_cache.py
• workflows/sbi/jraph_sbi_flowjax.py
• workflows/sbi/plot_flowjax_posteriors.py
• shared/eigenvalue_transformations.py
• shared/tng_pipeline_paths.py cache suffix behavior

Knowledge gaps addressed

• Existing docs emphasized only the default softplus/transformed cache path while recent wedge NPE work uses explicit linear increments.
• The cache suffix/mode coupling was implicit and easy to mismatch.
• Scale-invariant feature normalization and its transfer-use constraint were not discoverable from workflow docs.
• Resume/checkpoint and posterior class-probability diagnostics were implemented but under-documented.

Validation

• Confirmed documented flags/suffixes against source with ripgrep.
• Ran python3 -m py_compile shared/eigenvalue_transformations.py.
• Environment-dependent validation blocked: this Cloud machine has neither conda nor python on PATH, so cosmic_env help checks and python -m unittest discover -s tests/phase4 could not be run here.
• Graphify refresh also blocked because graphify is not installed in this environment.