feat: add required_features() and missing-data validation

[marcbal77]

Summary

Adds a public, stable required_features() method to every model plus a generic input-validation step so clocks raise a clear, actionable error when a GeoData lacks the markers or metadata they need. This is the consumer side of the clinical-clock work, split out of #203 per review.

API (stable at 1.0)

New biolearn.features module:

• RequiredFeatures(layer, features, metadata): a frozen dataclass that is also a Mapping, so it behaves as the documented {"layer", "features", "metadata"} dict and as an object with .layer / .features / .metadata. All three keys are always present.
• MissingFeaturesError(ValueError): raised on missing inputs. Subclasses ValueError, so existing except ValueError handlers keep working while callers can catch this specific case.
• validate_required_features(model, geo_data): the single validator.

Every model gains required_features():

gallery.get("Horvathv1").required_features()
# {"layer": "dnam", "features": [...], "metadata": []}

Validation scope

Hard-fail validation is wired only where a missing feature corrupts the result: LinearMethylationModel (the exact linear combination, which previously carried its own private check). Clocks that tolerate partial inputs by design keep their current behavior and expose required_features() for introspection only: PCLinearTransformationModel (intersect and center-fill), LinearTranscriptomicModel (inner join), MiAge, EpiTOC2, AltumAge, Deconvolution, GPAge, and the proteomic clock. GrimageModel and HurdleAPIModel keep their existing bespoke checks.

The validator encodes the one axis asymmetry between layers in a single place (_LAYER_FRAME): dnam / rna / protein are features-as-rows (features in .index); clinical is samples-as-rows (biomarkers in .columns), matching metadata.

Tests

New test_features.py:

• RequiredFeatures dict and object behavior; all three keys always present.
• The validator per layer (dnam index vs clinical columns), metadata checks, and backward-compatible ValueError.
• A regression test that a tolerant clock (EpiTOC2) still runs when a required CpG is missing.
• A gallery-wide smoke test that every model returns a valid RequiredFeatures, including forwarding through ImputationDecorator.

Full suite: 221 passed, 5 skipped.