[XiaoHongbo-Hope]

Purpose

Add an Arrow batch temporal matching kernel for PyPaimon alignment. It returns matching positions for grouped as-of joins, interpolation brackets and windows, avoiding one Python binary search per anchor row.

The kernel accepts sorted, non-null int64 keys and validated group bounds, releases the GIL while matching, and uses i128 arithmetic for timestamp limits. Duplicate selection, nearest ties, exact-match exclusion, tolerance and window closure follow PyPaimon's existing semantics. It performs no table reads or writes.

Validation

• Three standalone Rust kernel boundary tests pass.
• The full Rust extension plus PyPaimon integration tests: 101 passed, 155 subtests passed. Covers random differential cases, sliced arrays, empty/missing groups, duplicates, closure, int64 limits and invalid inputs.
• Full release extension build passes with Rust 1.94.0.
• Complete-output integration validation: all 300,000 rows checked across 500 synthetic episodes.

Consumer integration: apache/paimon#10846.

[JingsongLi]

The 18 temporal tests passed against an isolated build of the unchanged temporal binding. I found one Arrow input-handling issue; the inline comment includes a reproducer and a suggested fix.

[JingsongLi]

[P2] Align foreign buffers before constructing Int64Array

ArrayData::from_pyarrow_bound() preserves the foreign buffer address, but Int64Array::from(data) requires 8-byte alignment and panics when it is missing. PyArrow permits valid, sorted, non-null int64 arrays backed by unaligned buffers, including zero-copy NumPy inputs. I reproduced a PanicException with the unchanged temporal binding using Arrow 58.4.0:

import struct
import pyarrow as pa
from pypaimon_rust import temporal_match

raw = b'x' + struct.pack('<3q', 1, 3, 5)
times = pa.Array.from_buffers(
    pa.int64(), 3, [None, pa.py_buffer(raw).slice(1)])
times.validate(full=True)  # Valid, sorted, non-null int64 input
temporal_match(
    times, pa.array([1], type=pa.int64()),
    pa.array([0], type=pa.int64()),
    pa.array([3], type=pa.int64()), 'nearest')

This should return first=[0], second=[-1], but instead raises PanicException: Memory pointer from external source ... is not aligned with the specified scalar type.

Please call data.align_buffers() before constructing the typed array, as the existing arrow_input::array_from_pyarrow helper already does, and add a regression test for unaligned input. The same helper imports all four input arrays, so this affects more than just times.