fix: verify registered models against the stored spec on deserialization

[rahul188]

Problem

SignalSchema._deserialize_custom_type resolves custom types by name@version alone:

if fr := ModelStore.get(model_name, target_version):
    return fr

Since ModelStore is process-global and all generated models are @v1, two datasets whose inferred models share a name (both explode()d from a column named json, two parquet files with an items list-of-struct column, …) shadow each other: whichever model registers first is used to read every same-named dataset for the rest of the process. Depending on how the shapes overlap that surfaces as a KeyError (missing DB column), a Pydantic ValidationError, or a silent misread. Repro from #1872:

c1 = dc.read_values(json=[{"a": 1}]).explode("json")
c2 = dc.read_values(json=[{"b": "x"}]).explode("json")
c1.save("e1"); c2.save("e2")
dc.read_dataset("e2").to_values("json_expl")  # works
dc.read_dataset("e1").to_values("json_expl")  # KeyError: 'json_expl__b'

Fix

Following the approach sketched in the issue:

• A registered model resolves a custom type only when its field spec matches the spec stored in the dataset schema. The check reserializes the registered model with the existing _serialize_custom_model machinery and compares fields per custom type, so a divergence in a nested model (identical top-level field strings, different Inner@v1 shapes) is caught too.
• On mismatch, the model is recreated from the stored spec under a spec-fingerprinted alias ({name}_{fingerprint[:10]}, the same shape to_partial uses for partials) instead of clobbering the registered model's name in ModelStore. Repeated reads of the same dataset hit the alias cache and reuse the recreated class.
• When there is no registered model, behavior is unchanged (recreate under the stored name and register).

The fingerprint covers the transitive closure of the type's spec (fields, bases, hidden fields of every reachable custom type), so two colliding specs that differ only in a nested type or base still get distinct aliases.

Testing

• test_deserialize_same_name_different_shape_not_shadowed — the cross-dataset collision, plus re-read identity (same class object on repeated deserialization of both specs)
• test_deserialize_nested_shape_divergence_not_shadowed — identical top-level specs, diverging nested model
• test_deserialize_prefers_registered_model_when_spec_matches — round-trip still returns the registered class itself, not a recreation
• test_explode_same_column_name_across_datasets — the issue's end-to-end repro (explode → save → read_dataset both ways)

All four were run against unmodified main first: the three collision tests fail there (the e2e test with KeyError: 'json_expl__a'), and the registered-model test passes on both, confirming no regression of round-trip identity. Full tests/unit/lib/test_signal_schema.py (120) and tests/unit/lib/test_datachain.py (349) pass; ruff check, ruff format --check, and mypy are clean on the changed files.

Fixes #1872

[rahul188]

Hi 👋 Gentle nudge on this PR — it's mergeable and CI is passing, just waiting on a review. Would really appreciate a look whenever you have bandwidth, and I'm glad to make any changes you'd like. Thanks for your time and for maintaining this project! 🙏