[codex] add predefined RowBinary encoders

[ruslandoga]

Summary

Adds Ch.RowBinary.define_encoder/1 for compile-time RowBinary encoders from a non-Ecto schema shape such as schema: [id: "UInt64", text: "String"] or schema: %{"id" => "UInt64"}.

The generated encoder can:

• encode one atom-keyed or string-keyed row map/struct
• encode many rows with rows: true
• encode a full INSERT INTO ... FORMAT RowBinaryWithNamesAndTypes body when table: is provided, with the names/types prefix precomputed

The first specialization covers common hot insert paths and falls back to Ch.RowBinary.encode/2 for less common types.

Benchmark

Updated bench/encode.exs to compare current list/map paths against the generated encoder. Local Apple M2 result for 1,000,000 rows:

Predefined atom map encoder stream of 100k row chunks: 2.22 ips / 450.21 ms
Predefined atom map encoder: 1.96 ips / 509.09 ms
Predefined full insert body: 1.92 ips / 521.85 ms
RowBinary row lists: 1.00 ips / 1004.43 ms
RowBinary atom maps via fields: 0.81 ips / 1229.04 ms

Follow-up Issues

• Optimize RowBinary DateTime encoding in generated encoders #371 Optimize RowBinary DateTime encoding in generated encoders
• Explore column reordering and fixed-width packing for generated RowBinary inserts #372 Explore column reordering and fixed-width packing for generated RowBinary inserts
• Specialize more generated RowBinary encoder type paths #373 Specialize more generated RowBinary encoder type paths

Validation

• mix compile --warnings-as-errors
• MIX_ENV=test mix compile --warnings-as-errors
• MIX_ENV=test mix format
• git diff --check
• mix run bench/encode.exs
• custom mix run -e smoke check comparing generated encoder bytes against encode_row/2, encode_rows/2, and encode_names_and_types/2

mix test test/ch/row_binary_test.exs could not run in this workspace because the Docker daemon/OrbStack socket is unavailable, so ClickHouse from compose.yml could not be started.