Add STRUCT and LIST support for Arrow/JSON, fix DuckDB Arrow coercion

[texodus]

Nested struct/list ingest for Arrow and JSON

perspective-server now accepts Arrow LIST and STRUCT columns, as well as JSON/NDJSON Array and Object columns, via column expansion. This change makes it much easier to wire existing Arrow and JSON feeds from e.g. Kafka or csp directly into a Perspective Table without marshalling and transforming the input (and substantially reducing column copies when the transfer encoding is Arrow, since Perspective can route this directly).

• struct / object → one dotted column per leaf, recursively. A null parent
  nulls every descendant (e.g. struct itself is not nullable).
• list / array → controlled by a new list_flatten table option:
  • zip (default) — one row per element, non-list siblings repeated; an
    empty list yields one null row; multiple lists in a record must share a
    length.
  • cartesian — cartesian product.
  • stringify — JSON.stringify.

Virtual server window aggregates and type coercion

Features::window_aggregates carried a bare list of names per column type, which forced the SQL translation to know each function's shape. It now encodes these as generic WindowAggSpec, mirroring the existing declared AggSpec, which allows the model to specify:

• frames — which frame kinds the function accepts (rows, range,
  cumulative, none).
• offset — takes a row offset.
• alpha — takes a smoothing factor, as ema does.
• result_type — the output type.

duckdb.py and clickhouse.py virtual servers have updated their window aggregate catalogs against it, including ranking functions that take no source column and derived ones like rate. New backends should be able to add arbitrary window aggregates via the descriptor language.

Arrow type coercion

Completes the DuckDB/Arrow type coercion work started in #3151 with the additional missing types (sans arrow-compute). Add support for DECIMAL, TINYINT, ENUM, the unsigned types and timezone-carrying timestamps at any unit, fixes #3149

EDIT: This includes timestamp(μs, "UTC") which also has tests