feat: opt-in EXECUTE_ARRAY prepared batch path for executemany() #839

Problem and evidence

#836 measured a prototype EXECUTE_ARRAY (function code 21) path: one PREPARE, one request with every row's bound values, one CLOSE_REQ. At 10,000 rows, commit included, it takes 0.33-0.55x of the shipped executemany() for INSERT, UPDATE and DELETE and 0.15-0.16x for MERGE on CUBRID 11.0-11.4 (0.51-0.61x on 10.2). It is also 0.41-0.54x of the prepare-once loop. Experiment and raw data: cubrid-lab/cubrid-benchmark#23 (experiments/pycubrid-prepared-batch).

On 10.2, 11.0, 11.2 and 11.4 the server keeps the shipped batch's semantics:

• every row is executed and only the failing row is undone;
• autocommit commits each successful row;
• per-row counts and the rowcount sum are identical.

The path differs in three ways:

• it accepts only the prepared type set (feat: widen the prepared parameter types (BIGINT, DOUBLE, NUMERIC, temporal, binary) #840 widens it);
• values are coerced as host variables, not literals;
• with autocommit on, a LOB handle fails after the first row's commit (-1016).

So the decision in #836 is an opt-in path. The default stays the same.

Scope

1. Add ExecuteArrayPacket (request: query handle, query timeout for protocol V4+, autocommit byte, one type/value pair per marker per row; reply: per-row count + OID or error code + message, shard id for V5+), with offline packet tests.
2. Add an opt-in to executemany() on the sync and async cursors, for example a keyword-only prepared=True. Agree on the API in the design review first. For INSERT/UPDATE/DELETE/MERGE it runs PREPARE + EXECUTE_ARRAY, and it keeps the existing session-generation fence, the deferred-close handling (perf: release unclosed cursor handles in autocommit mode with a deferred close queue #488) and the reconnect rules.
3. Keep the batch semantics: raise the first returned error, as today, with the same exception mapping, and set rowcount to the sum of the per-row counts.
4. Reject values outside the prepared type set before anything is sent. Reject LOB handles when autocommit is on.
5. Add integration tests on 10.2-11.4 for duplicate-key and conversion errors in the middle of a batch (autocommit on and off), rowcount, NULLs and SET values, matching the chore(performance): evaluate a prepared-statement batch path for executemany #836 probes. Update PERFORMANCE.md, API_REFERENCE.md, PARAMETER_BINDING.md, the Korean copies, CHANGELOG.md and RELEASE_POLICY.md (additive API).

Non-goals

• Changing the default executemany() path.
• Server-side protocol changes.

Links

• Decision: chore(performance): evaluate a prepared-statement batch path for executemany #836
• Data: experiment: pycubrid prepared batch paths, including EXECUTE_ARRAY, on CUBRID 10.2-11.4 cubrid-benchmark#23