Add the READ operation and a snapshot reader for existing rows

[lukashes]

Second PR toward the initial snapshot (#49). Groundwork only: the reader is not wired into startup yet.

Problem

Replication starts at the slot's LSN, so a new consumer never sees rows that existed before the slot was created. #49 asks for those rows to be emitted once, up front, consistent with where the stream then begins.

Solution

• Add READ to ChangeOperation and "read" to the allowed stream operations. A READ event reuses the insert payload, so serialization, matchStreams, and partitioning are unchanged; a stream opts into the snapshot by listing read in its operations.
• New SnapshotReader (source/postgres/snapshot.zig): on a regular connection it enters a REPEATABLE READ transaction bound to an exported snapshot (SET TRANSACTION SNAPSHOT), reads each table with a cursor + FETCH to bound memory, and emits each row as a READ event stamped with the slot's consistent point as meta.lsn.
• Make converter.mapValue public so a READ row and a streamed change of the same column map (oid, text) to the same JSON type.

Tests

• Integration: exports a snapshot with pg_export_snapshot(), then checks the reader returns the pre-export rows only (a row inserted after export is not read) and that values are typed like the streamed path (int/bool/float). Plus an empty-table case.
• Unit: READ serializes to "op":"READ", read passes config validation, and a pure matchStreams check that a READ event routes only to streams listing read.

Startup wiring, the snapshot.mode gate, and e2e come in the next PR.

[github-actions]

📊 Benchmark Results

Current run is the minimum over 3 passes, compared against the base branch (main @ 8fda403), built on the same runner.

Benchmark	Baseline	Current	Δ Time	Allocs	Status
Converter DELETE	141.90μs	142.13μs	+0.2%	13 → 13	➡️
Converter INSERT	141.81μs	142.41μs	+0.4%	13 → 13	➡️
Converter UPDATE	171.21μs	171.41μs	+0.1%	22 → 22	➡️
JsonSerializer	50.21μs	49.77μs	-0.9%	3 → 3	➡️
KafkaProducer produce	215.05μs	218.20μs	+1.5%	1 → 1	➡️
KafkaProducer send	253.90μs	278.15μs	+9.5%	0 → 0	🔴 slower
PgOutputDecoder	95.48μs	94.88μs	-0.6%	6 → 6	➡️
getPartitionKeyValue boolean	15.40μs	15.35μs	-0.3%	1 → 1	➡️
getPartitionKeyValue integer	0.04μs	0.04μs	+5.0%	0 → 0	⚪ noise
getPartitionKeyValue not found	0.03μs	0.03μs	+3.1%	0 → 0	⚪ noise
getPartitionKeyValue string	17.23μs	17.10μs	-0.7%	1 → 1	➡️
matchStreams found	17.43μs	17.20μs	-1.3%	1 → 1	➡️
matchStreams not found	0.06μs	0.05μs	-1.8%	0 → 0	⚪ noise

Summary: ➡️ 9 neutral · 🔴 1 slower · ⚪ 3 ignored (sub-μs)

_{Thresholds: <1μs ignore · 1–20μs 15% · 20–50μs 10% · ≥50μs 5%. Measured on a shared CI runner — treat small deltas as noise. Informational only; this check never fails the build.}