Adopt the core runtime and replace metadata-blind Zarr reads

[uermel]

Parent epic: #84

Implementation: #93, stacked on #94

Goal

Adopt the published copick 2.0 alpha runtime and replace metadata-blind tomogram reads with one explicitly read-only, metadata-based accessor.

Runtime metadata

• Python >=3.11
• Zarr >=3.1.6,<4
• copick>=2.0.0a1
• pydantic>=2, with no upper bound

Pydantic is direct because copick-utils imports BaseModel. The locked environment deliberately resolves Pydantic 2.13.4. Stable Python 3.14.7 passes; the earlier failure was an obsolete uv pin selecting Python 3.14.0b2, not a Pydantic incompatibility with stable Python.

Implemented scope

• Raise the Python runtime, update classifiers/package URLs, and regenerate the lock.
• Add copick_utils.io.zarr.get_level_array(entity, level=0) using zarr.open_group(..., mode="r") and copick.util.ome.get_level_path().
• Replace all eight audited direct tomogram reads.
• Treat level 0 as the first metadata dataset rather than literal path "0".
• Preserve lazy array access and fail clearly for invalid levels/metadata.
• Smoke-import all implementation modules.
• Load all 32 command entry points and render their help paths.
• Run the locked behavior suite on stable Python 3.11 through 3.14.

Tests do not assert dependency declaration strings or maintain a separate minimum-dependency job. The declared bounds are exercised by the locked clean installation and behavior matrix.

Acceptance criteria

• Numeric and nonnumeric OME dataset paths work for Zarr v2 and v3 fixtures.
• Invalid levels fail through the bounded core helper.
• Reads are explicitly read-only and preserve the source layout.
• All implementation modules and command entry points load.
• The locked stable-Python matrix passes.
• feat: resolve tomogram levels from OME metadata #93 passes and merges after test: establish Zarr migration baseline #94.