Investigate numpy-array-like Arrow view to avoid the scatter copy

[ghostiee-11]

SQLBackendArray._raw_getitem and the eager aggregation path both copy PyArrow RecordBatch columns into a preallocated numpy buffer (one read pass per batch column). The copy is necessary today because PyArrow buffers are immutable views and numpy needs a writable buffer.

xarray.Variable does not strictly require a numpy ndarray; it accepts anything that meets the _array_like contract (see xarray/core/variable.py#L373). If we can wrap a pyarrow.Array in a tiny adapter that satisfies that contract, the scatter step becomes a pointer assignment and the round-trip skips a full read pass.

Open questions:

• Do we need contiguous storage per dim, or can we hand xarray a list of pyarrow.Array chunks with a custom __array__ that materializes lazily?
• How does indexing (slicing, fancy indexing) compose with the adapter once xarray asks for sub-views?
• Does this play nicely with the chunked / dask path users will reach for after .chunk()?

Worth prototyping on a single dtype family first (float64) to see if the adapter shape is even tractable.

Spun out of #167.

[alxmrs]

I like the idea a lot, it's a cool researchy topic. Before starting, it would be worth doing some complexity analysis (or, writing micro memory benchmarks) to see what the potential win is here -- specifically, to see if the status quo is enough.