Well behaving streams only wait for some data to be available
(e.g. FTP, HTTP, raw TCP).
Poory behaving streams wait for the whole requested amount
(e.g. regular file - though it does not use prefetch anyway).
Badly behaving streams wait for the whole requested amount and do not
perform any pipelining/buffering (e.g. CIFS, probably SFTP).

In the best case, this change reduces the number of calls, thus
slightly improving performance.

In the worst case, it reduces the number of round-trips necessary,
thus restoring support for playback at bandwidth-latency product.
However it increases latency in the low bandwidth case, but there is
a simple work-around: capping the read size to a suitably small value.

Another work-around which works both ways is to provide pf_block rather
than pf_read. But it incurs an extra memory copy.