v0.36.1 — window rid docs, split_many fixture

Docs: a window aggregate's rid is chunk-relative

The per-row window callback's rid was documented as the output row's index within
the partition. It is the row's index within the current output chunk. DuckDB's
window operator evaluates one chunk at a time and passes i (0 ≤ i < 2048); the
partition row is never sent. The extension forwards that value untouched. A worker
that used rid to index the partition read the wrong rows after the first chunk.

Use the frames instead: they are always partition-relative. The batched path's
row_idx is partition-relative and was already documented correctly.

Updated in AggregateWindowRequest, AggregateFunction.window() and
AggregateClient.window(). Docstrings are outside the protocol hash, so no port's
reflection changes.

Fixtures

• split_many honours its splits argument, as every other SDK's twin does (it
  returned max(splits, 1000)). splits <= 0 still means 1000. With the matching
  change to more_splits_than_threads.test (64 splits), that file went from 13.0 s
  to 1.2 s over HTTP.