Increase performance of IOStream.read_until and read_until_regex.

Increase performance of IOStream.read_until and read_until_regex.

_handle_read and _try_inline_read now only call _read_from_buffer
once, after calling _read_to_buffer as many times as they can.  This
allows the progressive _double_prefix calls in _read_from_buffer to
work as efficiently as possible.

In testing with a 4MB read, performance improved by a factor of 32 (8
seconds to 0.25 seconds)