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.