[python] Add as-of joins for multimodal scans

[XiaoHongbo-Hope]

Purpose

Add a lazy binary as-of join for independently sampled multimodal streams.

The API follows pandas merge_asof, Polars join_asof, and DuckDB ASOF JOIN: preserve left rows, match at most one right row within by, and support backward, forward, or nearest matching with optional tolerance.

steps = join_asof(
    actions.scan(),
    images.scan().select("image"),
    on="event_time",
    by="episode_id",
    direction="nearest",
    tolerance=20,
).join_asof(
    states.scan().select("state"),
    direction="backward",
    tolerance=50,
)

Chained joins remain lazy. Paimon pins snapshots, plans compact metadata with _ROW_ID, fetches matched payloads by batch, and keeps BLOBs as descriptors.

Resampling, interpolation, windows, and streaming/watermark semantics are out of scope.

Tests

94 passed, 6 subtests passed