[XiaoHongbo-Hope]

Purpose

Temporal alignment plans each input and reads each source payload serially, adding their remote I/O latency. Add opt-in to_arrow_batch_reader(parallelism=4) to overlap independent source planning and per-batch payload reads while preserving input/output order. The default remains serial.

Source-level concurrency is bounded by the requested parallelism and source count. Each batch finishes before the next is consumed. On failure, queued futures are cancelled and workers skip source operations that have not started; running tasks are joined before propagating the error. Existing snapshot, authorization and projection paths remain in use. Per-table reader concurrency still applies, and custom catalog/auth callbacks must support concurrent calls when enabled.

Validation

• Temporal suite: 93 passed, 1 skipped, 176 subtests passed.
• Added checks for serial/parallel output equality, bounded source concurrency, independent planning, worker cleanup, and argument validation.
• Added a two-worker regression with a blocked first source, a failing second source, and a third source that must not execute, even if its wrapper is dequeued before caller-side cancellation.
• Flake8 and git diff --check passed.