feat(timeseries): propagate ILP-inferred schema to catalog for SQL column resolution

[emanzx]

Summary

• Propagate ILP-inferred schema to the catalog: On first ILP ingest, the Data Plane infers a columnar schema (tags → Symbol/VARCHAR, fields → Float64/Int64, timestamp). Previously this schema only existed in the in-memory memtable, invisible to the SQL planner which still saw the default (timestamp, value) fields from CREATE TIMESERIES. Now the inferred schema is returned in the ingest response and the ILP listener updates StoredCollection.fields in the catalog.

• Fix Arrow schema for timeseries collections: Omit the synthetic id column (timeseries rows don't have document IDs) and map TIMESTAMP to DataType::Int64 (internal millis representation) instead of Utf8.

Unblocks queries like:

SELECT qtype, COUNT(*) FROM metrics GROUP BY qtype;
SELECT client, AVG(elapsed_ms) FROM metrics GROUP BY client;
SELECT MIN(elapsed_ms), MAX(elapsed_ms) FROM metrics;

Test plan

• ILP ingest 1M rows from AdGuard Home DNS query logs at ~180K rows/sec
• DESCRIBE collection shows all ILP-inferred columns after first ingest
• GROUP BY on tag columns (qtype, client, status) returns correct counts
• AVG/MIN/MAX on numeric fields (elapsed_ms) returns correct values
• WHERE + GROUP BY works for aggregate queries
• Non-timeseries collections unaffected (schema still includes id prefix)
• Schema only sent on first ingest batch (no overhead on subsequent writes)
• cargo clippy clean, cargo fmt --all applied