Move the B-channel message handling into a per-span I/O thread,
to solve most of the problems caused by the intermixed data + control
socket interface of mISDN, missing write poll() support on
mISDN B-channels and the FreeTDM I/O model. This eliminates most of
the audio problems (except for a few minor glitches).

A unix stream socket pair is used as a bi-directional pipe replacement
(the pipe code is still included in this commit, but will be removed later),
with the RX and TX buffer sizes carefully tuned to avoid excessive buffering
(= latency) and a deadlock situation between the write() call in ftdm_write()
and the code in misdn_span_run() that needs a minimum amount of data in the
TX buffer, before sending out a PH_DATA_REQ to the mISDN socket
(see misdn_span_run() comments for more details).

The minimum size for pipes is PAGE_SIZE (4k), which is ~500 ms worth of
audio. A socket pair RX/TX buffer size of 3k, seems to hold a maximum
amount of around 500 bytes data in practice, giving us a much lower
maximum latency than a unix pipe. (The socket pair might be replaced by a
custom ring buffer / fifo data structure to get even more fine grained
control of the maximum latency.)

The newly introduced span_start / span_stop callbacks in
ftdm_io_interface_t are used to start / stop the I/O thread. The callback
functions will wait up to 10 seconds for the thread to successfully
start up or shut down (using a mutex + condition var).

NOTE: Using any of the locking ftdm_span_() functions in the I/O will cause
      a deadlock between the I/O thread (trying to lock span->mutex) and the
      thread calling ftdm_start()/_stop() (holding the span->mutex).
      (The I/O thread currently uses direct span member access to avoid this.)

The I/O thread uses the epoll(7) family of functions for event handling.
An epoll context is created on startup and all B-channel sockets are
registered (READ, PRI and ERR). Before entering the event loop,
the I/O thread will send a signal on the condition variable, to
indicate it has completed the startup procedure.

Incoming b-channel and command pipe events are handled by the event loop.
Payload of incoming PH_DATA_IND frames (= audio data) is sent to the
rx_audio_pipe_in end of the b-channel's socket pair and, if enough data is
available, a PH_DATA_REQ of the same size is sent to the b-channel mISDN socket
to transmit audio.

A MISDN_CMD_STOP command on the event pipe will wake up the I/O thread and
cause it to shut down. All b-channels will be unregistered from the epoll context
and the epoll fd closed. The I/O thread terminates itself after signalling the
successfull shutdown on the condition variable.

TODOs:
    - Move D-Channel into I/O thread too

    - Custom FIFO/ring buffer for data (even lower latency)

    - Improve epoll() code (per-channel struct w/ callback, for epfd.data.ptr)

    - Use mISDN DSP for audio (e.g. tone generator, dtmf detector, echo cancel)

    - Use a per-port / span control socket to execute channel commands
      synchronously, or add misdn_commands (queue?) that can be used that way

    - Name I/O threads 'mISDN-%SPAN_NAME%', e.g. 'mISDN-M_BRI1'
      (= add ftdm_thread_set_namef(thread, fmt, ...) / ftdm_thread_set_name(thread, name))

TL;DR: "tweak", solves "booboo" with audio
Signed-off-by: Stefan Knoblich <stkn@openisdn.net>