Add spatial_expression dataset_type with sidecar target-matrix plumbing

[clemsgrs]

Summary

A sanctioned prefactor that teaches soma's Manifest a multi-target regression supervision shape, so later HEST slices can curate and probe gene-expression targets without touching the schema again ("make the change easy, then make the easy change"). No curator and no probe here — only the schema shape and its load path.

Introduces a new dataset_type — spatial_expression — whose single supervision column is target_index, an integer row-key into two sidecars written beside dataset.csv:

• targets.npy — shape [n_rows, n_genes], row i is the target_index==i spot's vector.
• genes.json — the ordered gene list.

One sample = one spot. The loaded SampleRecord carries its resolved vector on target; SpatialExpressionManifest attaches it from the sidecar.

Acceptance criteria

• spatial_expression is a recognized dataset_type with target_index as its supervision column in the shared SUPERVISION_COLUMN map.
• The shared write_manifest emits a valid spatial_expression Manifest (dataset.csv with sample_id, image_path, target_index) plus the targets.npy and genes.json sidecars; the loader rejects a Manifest missing the sidecars or the column (fail-fast).
• Loading a hand-built spatial_expression Manifest yields sample records whose vector targets round-trip exactly (values and gene order) from the sidecar.
• Re-writing the same Manifest is byte-identical (deterministic).
• Existing dataset_types (tile/slide/patient/segmentation/detection) are unaffected.

Serialization choices

• Target matrix → targets.npy (uncompressed .npy via np.save, allow_pickle=False, C-contiguous). Uncompressed keeps re-writes bit-deterministic; np.savez_compressed (used elsewhere for lossy heatmaps) is not byte-stable across zlib versions. Dtype is preserved verbatim from the caller (the design bakes float64 log1p counts).
• Gene list → genes.json (json.dumps(list(genes), indent=2) — order preserved, never sorted, since gene order is load-bearing).
• Both sidecar paths are surfaced on CuratedManifest (target_matrix_path / genes_path, None for other dataset_types). Filename constants live in soma.dataset and are imported by the writer to keep writer/reader in lockstep.
• SampleRecord.target is compare=False (derived supervision data, not sample identity; array fields have no scalar truth value).

Validation added

Writer and loader both fail-fast on: a missing sidecar, a non-2D matrix, a gene/column-count mismatch, and out-of-range target_index. Passing target_matrix/genes for any non-spatial_expression dataset_type is rejected.

Deviations / notes

• PipelineConfig._valid_dataset_types (in soma/config.py) is intentionally not extended in this slice. There is no probe/task yet to make a spatial_expression pipeline runnable, and adding it would force premature aggregator/task constraints. The load path (load_manifest) — this issue's scope — recognizes it; the runnable-pipeline wiring lands with the probe (design §10, PR 2).
• No hest extra or HEST curation dependency added (deferred, per the issue).

Tests

Added 11 tests in tests/test_curation_manifest.py covering the supervision-column map, sidecar emission + round-trip, byte-identical re-write, loader dispatch/round-trip, missing-sidecar / missing-column rejection, and the writer's range/count/misuse guards.

Full suite: python -m pytest -q → 1513 passed, 3 skipped, 1 xfailed.

Closes #257