Multiple threads iterating over a file can corrupt the file's internal readahead
buffer resulting in crashes. To fix this, cache buffer state thread-locally for
the duration of a file_iternext call and only update the file's internal state
after reading completes.

No attempt is made to define or provide "reasonable" semantics for iterating
over a file on multiple threads. (Non-crashing) races are still
present. Duplicated, corrupt, and missing data will happen.

This was originally fixed by 6401e567, which
raised an exception from seek() and next() when concurrent operations were
detected. Alas, this simpler solution breaks legitimate use cases such as
capturing the standard streams when multiple threads are logging.