Replace the Arrow scan with a port of the neo branch's scan

[evertlammerts]

Registered Arrow objects are wrapped in a pyarrow Scanner that evaluates pushed filters with pyarrow kernels on every batch and reads one shot streams far ahead of the query. That caused the kernel failures of #663 to #666, the memory growth of #691, wrong rows for struct columns in sliced batches, and a hang at exit with an unread generator reader.

This PR replaces that path with a port of the neo branch's scan: one table function, python_arrow_scan, over a private source adapter.

• The engine offers each filter to the source, which takes only what it applies exactly (a pyarrow expression on Datasets, a polars one on LazyFrames); the rest stays above the scan. In memory objects never receive filters.
• Only the columns a query reads are converted. Every array is validated before adoption; a stream whose schema differs from the declared one is refused.
• Streams are read one batch per thread, without read ahead and without holding the GIL; a one shot reader read twice gets a clear error.
• pa.Array, ChunkedArray, pandas and polars Series and polars LazyFrames are accepted everywhere; every Python error from a source is an InvalidInputException naming the object.
• Interpreter exit waits for a pull inside Python and refuses new scans.

Notes:

• No longer pushed, the engine filters instead: LIKE, NaN, nanosecond and infinite constants, date64 columns, and float predicates NaN can satisfy on Datasets and LazyFrames (parquet row group statistics leave NaN out). In memory row counts are estimates.
• A struct below a struct in a sliced batch is refused until the engine folds the parent offset; it read wrong rows before.