[Python] Throw an error if a RecordBatchReader is read more than once

[Tishj]

Since PyArrow RecordBatchReader objects are destructive (because they are stateful), they don't behave the same as native DuckDB tables, or Pandas/Polars DataFrames.

When we detect that a RecordBatchReader is encountered twice in a replacement scan, we throw an error.

Remaining Issues

State is Connection-local

This state is kept inside the PythonContextState, which is created for every separate connection, meaning that this error does not occur when two separate connections read from the same record batch reader.
We should instead make the RecordBatchReaderRegistry global on the module level, using locks to ensure multiple connections can make use of it at the same time.

Canceled Relations

We record the RecordBatchReader the moment it is encountered in the replacement scan, relations could theoretically be canceled, so they never run.
Subsequent relations that reference the RecordBatchReader then cause an error to be thrown.
We should differentiate between found and consumed RecordBatchReaders so this doesn't happen.

Another somewhat related issue to "Canceled Relations", using a LIMIT it could be possible to read from the same record batch reader and have it not be an error, if queries are issued to partially read from the RecordBatchReader