Projection pushdown into every decoder

Applies to: the JSON, Avro, and protobuf decoders

What it is

The query's projection narrows what each decoder builds, rather than decoding every wire field and discarding the unread ones afterward (64ddc2a, 83b3d69, 86908f1, 4af9d63).

• JSON decodes straight to the narrowed schema.
• Avro keeps the full writer schema (required for correct decoding) but materializes only the reader-schema columns.
• Protobuf prunes its descriptor so unread fields are skipped on the wire.

Measured, and why the split

Profiling drove the split: build/copy-bound formats gain a lot from pruning, tokenize-bound JSON gains little, because JSON's cost is the parse itself rather than building/copying values.

• Avro: Kafka/Avro q0–q2 1.06–1.18x → 1.64–1.83x.
• Protobuf: 0.88–0.94x → 1.26–1.36x.