fix(copy): respect auto_checkpoint setting on COPY FROM (#755)

[adsharma]

Fixes #755

COPY FROM unconditionally forced a checkpoint, ignoring DBConfig::autoCheckpoint (the CALL auto_checkpoint=... setting) and DBConfig::checkpointThreshold. The forced branch in transaction_manager.cpp:169 ran first, so the auto-checkpoint branch never got a chance. The result was a large per-statement cost on every COPY that the user could not turn off.

Change (src/main/client_context.cpp, 1-line guard):

if (preparedStatement->getStatementType() == StatementType::COPY_FROM &&
    getDBConfig()->autoCheckpoint) {
    // Note: We force checkpoint for COPY_FROM statement unless the user has
    // disabled automatic checkpointing.
    Transaction::Get(*this)->setForceCheckpoint();
}

getDBConfig() is already in scope; autoCheckpoint is the boolean behind the CALL auto_checkpoint=... setting. The default behaviour is unchanged. When auto_checkpoint=false, COPY still succeeds but no longer triggers a checkpoint, so the data lives in the WAL until the next auto-checkpoint or a manual CALL db_checkpoint();.

Test (test/test_files/transaction/copy_no_auto_checkpoint.test):

• CopyNoAutoCheckpointSmokeTest — CALL auto_checkpoint=false; then COPY foo FROM (UNWIND ...), asserts success and that the setting is preserved. This is enough to detect a regression of the original bug.
• CopyWithDefaultAutoCheckpoint — default auto_checkpoint=true, COPY succeeds.

Test would benefit from a stronger WAL-visibility assertion once a wal_replay_info-style public API is available; none exists today.

Test results: new tests pass; all 67 transaction tests continue to pass.

Perf note: a best-effort perf profile against the tinysnb COPY test shows the runtime is dominated by CSV parsing (regex_traits::transform, collate::do_transform, ~77%), not the COPY kernel. The dataset is too small to draw conclusions about the non-checkpoint COPY overhead. A larger benchmark would be the right next step for the part-2 measurement in the issue.