fix(bind): honour the arrow.json tag through dictionary encoding; test null-in-dictionary-values

[fornwall]

Two follow-ups from reviewing fornwall/arrow-adbc#34 (dictionary bind support in the C SQLite/PostgreSQL drivers) against this driver. Our bind.rs already uses the delegate-to-the-plain-path design that PR retrofits — the key selects the dictionary value, which re-enters the same cell_value mapping — so its whitelist-drift bugs can't occur here by construction. The comparison surfaced one genuine gap and one untested-but-correct edge case:

JSON through dictionary encoding

is_json_field required the field's own data type to be utf8-family, but the Arrow spec allows an extension array to be dictionary-encoded (the field's storage type is then dictionary<indices, utf8>). So a JSON-tagged Dictionary(Int32, Utf8) column — e.g. a pandas categorical of JSON strings — silently drifted from its plain form:

• it bound as plain STRING, which Spanner refuses to coerce into a JSON column, and
• ingest create modes made a STRING(MAX) column instead of JSON.

is_json_field now looks through Dictionary(_, storage) for the storage check. Present cells then get the explicit JSON param type for free via the existing plain-path delegation; null_dictionary_value now takes the field so a null cell keeps the type too (the plain path's typed-null rule). The tag stays ignored on non-string dictionary values (Dictionary(_, Int64)), matching the plain-storage rule.

Regression test: non-null key → null dictionary value

A dictionary cell can be null two ways: a null key, or a valid key selecting a null entry inside the values array. All existing dictionary tests only exercised the first. The second is handled correctly today, but only as a side effect of the delegation design (the delegated scalar binder's is_null check fires on the values array) — and it is exactly the case that stayed latent for years in the ADBC postgres driver's already-supported string dictionaries (insufficient data left in message, found in fornwall/arrow-adbc#34). The new test locks it in so a future restructuring of the dictionary arm fails the suite instead of reintroducing that bug.

No wire-format or option changes; cargo test, clippy -D warnings and fmt --check are clean.

🤖 Generated with Claude Code

https://claude.ai/code/session_01BgsLAmXUPFThybc8nfF4Eh