feat: clinical layer on GeoData + load NHANES/FHS through the library

[marcbal77]

Summary

Simple foundation for clinical data in biolearn:

• GeoData.clinical layer (samples-as-rows, biomarkers-as-columns — same orientation as metadata, matches industry convention for tabular clinical data)
• load_nhanes_as_geodata(year) and load_fhs_as_geodata() so both data sources flow through the library and return a GeoData
• A biomarker registry with unit conversions, validated end-to-end against real FHS Period 1 data (glucose mg/dL to mmol/L)
• An example (examples/01_composite_biomarkers/plot_load_nhanes_through_library.py) showing the through-the-library pattern

What changed since the first round

Based on your review:

• Orientation: Clinical data is now samples-as-rows (industry standard). Each row is an entity (patient), each column is a thing about that entity (biomarker). Matches metadata.
• required_features(): Dropped entirely from this PR. The schema-check / missing-data consumer path is a separate PR.
• UK Biobank preset: Removed. We have not validated the library against real UK Biobank data, so claiming support would be misleading. Swapped in an fhs preset that is validated end-to-end against the real Framingham Heart Study Period 1 data we already load.
• Canonical units: Fixed albumin (g/L) and creatinine (umol/L) to match what load_nhanes actually returns from the SI columns.

Test plan

• make test: 198 passed, 5 skipped
• make format: clean
• FHS source preset validated end-to-end against the real frmgham2.csv (test_load_fhs_as_geodata_applies_fhs_unit_conversion)
• load_fhs and load_fhs_as_geodata produce matching glucose values
• Save/load roundtrip with the clinical layer
• All 69 existing clocks still pass via the unchanged test_model.py

[sarudak]

Should add a data library entry for this.

[marcbal77]

Done. NHANES and FHS are now registered as DataLibrary entries, loaded via the same pattern as the GEO sources:

from biolearn.data_library import DataLibrary
data = DataLibrary().get("NHANES_2010").load()

Added NhanesParser and FhsParser classes in data_library.py, three YAML entries (NHANES_2010, NHANES_2012, FHS), and removed the top-level load_nhanes_as_geodata and load_fhs_as_geodata convenience functions so there's a single path. The example uses DataLibrary too.

[sarudak]

LGTM

Not sure if clinical or bloodwork is a better property name

[marcbal77]

Good point. Went with clinical since the plan has a few clocks that pull in non-blood inputs too (BMI, BP, body composition for LinAge2, CardioMetAge, Biometric Aging). Clinical keeps the layer accurate for all of them. Happy to revisit if that ever narrows to blood only.