[Load] Preserve pending tablets when parser fails during fallback

[Caideyipi]

Problem

When tree-model Load falls back from scan parsing to query parsing, a parser exception can occur after earlier Tablets have been converted but before the pending batch reaches its normal flush boundary. The exception path released the conversion context and discarded that batch.

Fix

Flush pending Tablets before returning a normal parser error. Temporary-unavailable results retain the pending batch and original parser status for the next managed-task retry; memory-pressure and interruption exceptions keep their existing retry behavior.

Tests

• mvn -pl iotdb-core/datanode -am test -Dtest=LoadTreeStatementDataTypeConvertExecutionVisitorTest -Dsurefire.failIfNoSpecifiedTests=false -DfailIfNoTests=false -DskipITs
• Spotless check
• Checkstyle: 0 violations

The separate exact-query versus wildcard-query observation is intentionally out of scope.

[luoluoyuyu]

The idea is reasonable: flush pending tablets before reporting a parser failure so already-converted data isn’t lost. Two things need confirmation though: 1) In the synchronous pipe path, a non-success result is marked retryable by the caller, but this change clears the conversion context after flushing, so a retry can restart from the beginning and re-insert the same tablets; 2) this also affects ordinary LOAD, where a parser failure would now become partial write + error. Consider limiting the flush to receiver-owned/terminal tasks, or mark the task terminal after flushing, and add coverage for the managed path and for retries not duplicating writes.