[Python] Ensure (Chunked)Array/RecordBatch/Table methods don't crash with non-CPU data

[jorisvandenbossche]

Currently, many of the methods of pyarrow Array/ChunkedArray/RecordBatch/Table crash when the object is backed by buffers pointing to non-CPU data (i.e. all methods that actually do something with the data and not just return info like the schema or the length).

We should ensure users don't see segfaults but get a proper error about the method only being support for data on the CPU. Probably that will require adding some "ensure cpu" helper function to most methods in the Python bindings.
This helper function will then have to inspect the device of the buffers and ensure all of them are on the CPU (not sure if there already exists a helper for this on the C++ side?)

• [Python] Array fails gracefully on non-cpu devices #42112
• [Python] ChunkedArray should fail gracefully on non-cpu devices  #43728
• [Python] RecordBatch should fail gracefully on non-cpu devices #43727
• [Python] Table should fail gracefully on non-cpu devices #43973

[jorisvandenbossche]

This helper function will then have to inspect the device of the buffers and ensure all of them are on the CPU (not sure if there already exists a helper for this on the C++ side?)

In bridge.cc, we already have ValidateDeviceInfo which loops recursively over all buffers of the array and its children, but that is C Data interface specific (the error messages). Similar logic can be added to the main ArrayData code as well, which is already being done in #40807 (which is adding a device_type property on ArrayData)

[jorisvandenbossche]

Closing this issue, with all subtasks being completed.

Another related issue, to do the same for compute functions, is still open -> #43541