feat(core): chDB connection pool, Arrow WAL zero-copy flush, and InfluxQL parity

[austin-barrington]

Re-architect the write path around chDB-ready Arrow batches, enable real
same-path connection pooling for concurrent flush inserts and queries, and
close major gaps in InfluxDB v1 TimeseriesQL semantics (DDL, SHOW, CQ
scheduling, rollups/MVs, SELECT INTO). Bump workspace to 0.8.3.

This is a cross-cutting release: ingest, WAL, flush, chDB adapter, query
translation, cluster schema apply, and observability all move together so
the fast path is correct end-to-end rather than bolted on in one layer.

• Add ChdbConnectionPool: N independent Connections to the same
  --path, each with its own ChdbClient mutex (libchdb process-global
  singleton per path — see chdb/insert/concurrency.rs).

• Round-robin checkout with try_lock on busy slots; clamp pool_size
  to 1..=32 (default 4).

• Rewire ChdbNativeAdapter, ChdbQueryAdapter, and ChdbSession to
  use the pool instead of a single shared session.

• Fix config/docs: chdb.pool_size > 1 is now real parallelism, not a
  deprecated no-op. Recommend server.max_concurrent_queries >= pool_size.

• Update system-architecture.md to describe same-path multi-connection
  semantics (replacing the old per-slot subdirectory model).

• Build chDB-ready fact-table RecordBatches at ingest time
  (application/arrow_ingest/ for line protocol, columnar, msgpack, points).

• Introduce PreparedWalSlot / PreparedMeasurementBatch domain types with
  post-assign ingest_seq patching (domain/prepared_wal.rs).

• Coalesce sparse prepared batches per measurement (domain/arrow_coalesce).

• In-memory WalArrowCache indexes unflushed prepared slots by sequence;
  bounded take_range(from, to_inclusive) avoids evicting post-snapshot
  entries (fixes steady ~50% cache-miss under continuous load).

• Versioned on-disk WAL encoding via wal_ipc (HBWA magic, v1): optional
  storage.wal_format = "arrow_ipc" alongside legacy bincode.

• FlushService flushes prepared chunks directly through
  insert_record_batch_direct — no re-parse/re-coalesce on the hot path.

• flush.arrow_wal_enabled (default true) gates the RAM cache; metrics
  gauge hyperbytedb_wal_arrow_cache_entries for growth/OOM watch.

• application/wal_append.rs bundles prepared slots with legacy WalEntry
  for peer sync compatibility.

• build_prepared_wal_slot, write_prepared_batch, schema cache refresh
  from metadata, and field-type widening reconciliation
  (ALTER TABLE ... MODIFY COLUMN when metadata union exceeds cached types).

• Engine DDL: raw facts use ReplacingMergeTree(ingest_seq); rollup/MV
  destinations with additive partials use SummingMergeTree on sum columns.

• Pad sparse legacy WAL batches to full ensured column sets before insert.

• New HTTP /internal/chdb adapter hook for admin/debug paths.

• Depend on chdb-rust feat_arrow_insert (Arrow C Data Interface insert);
  Docker builds clone that branch; root/proxy Dockerfiles stub all workspace
  crates for layer caching.

• Split parser: dedicated lexer.rs + ddl_parser.rs for token-driven
  InfluxQL DDL/SHOW/auth (CREATE/DROP/ALTER DB/RP/user, GRANT/REVOKE,
  SHOW DATABASES/MEASUREMENTS/TAG KEYS/TAG VALUES/FIELD KEYS/SERIES/CQs/MVs).

• Major to_clickhouse.rs expansion:

  • Raw selects always project time, ascending order.
  • GROUP BY time defaults, fill/null/with bounds, tag ordering.
  • Materialized view backfill column ordering and dest insert mapping.
  • Rollup fact views (build_coalesced_fact_view_*): sum for additive
    fields, mean → sum/count rewrite on rollup measurements.
  • CQ bounded SELECT INTO translation and time-window predicate stripping.

• predicate_sql: shared WHERE → SQL for DELETE / DROP SERIES (local +
  replication).

• field_type domain module; rollup combine semantics for MV/CQ fields.

• InfluxDB v1 CQ scheduling (domain/cq_schedule.rs): bucket alignment,
  RESAMPLE EVERY/FOR validation, coverage windows, boundary-aligned
  should_run, execution interval derivation.

• QueryService::execute_continuous_query; reconstruct CQ text for replay.

• MaterializedViewService and ContinuousQueryService wired to new
  schedule metadata and bounded backfill paths.

• Peer/cluster: PeerQueryService Raft mutation forwarding, leader addr
  resolution (forward node → Raft → cluster membership → metrics leader),
  MV source/dest retention policy resolution.

• schema_mutation_apply: single apply path for Raft state machine,
  /internal/replicate-mutation, and startup metadata sync (metadata +
  chDB DDL side effects).

• RocksDB metadata adapter extended for CQ schedule fields, rollups, and
  richer measurement meta.

• Replication apply, hinted handoff, drain, bootstrap, and Raft log/state
  machine updated for prepared WAL and schema mutations.

• Expanded SHOW/DDL execution, SELECT INTO, retention policy normalization,
  tag key/value discovery from series tables, authorization checks.

• CLI 0.8.3: admin/query/export/import/repl hooks for new statement types;
  e2e test coverage extended.

• tikv-jemallocator with background purging: return transient startup
  heap (series dedup warm + WAL replay) to the OS instead of pinning RSS.

• Default retention sweep interval 12h (was 60s).

• Grafana dashboards refreshed (cluster, logging, machine-monitoring);
  Kind CR manifest and docker-compose aligned with new config knobs.

• scripts/load.sh updated for pool/Arrow WAL load testing.

• New compat suites: combination_tests (full parse→translate→execute
  interaction tests), cq_tests, prepared_wal_tests.

• Expanded ddl_tests, query_tests, metadata_tests, http_tests.

• Integration/raft/sync_quorum tests updated for prepared WAL and pooling.

• Bench stubs adjusted for new ingest signatures.

BREAKING CHANGE: chDB session pooling semantics changed — pool_size now
opens multiple same-path connections (real concurrency) instead of being
ignored/warned. Tune pool_size and max_concurrent_queries together.
New config keys: storage.wal_format, flush.arrow_wal_enabled.
Default retention interval is now 12h.