natpmp: fix out-of-order responses bug

This patch fixes a few issues with the way libnatpmp was used in DHTNet.
The two most important ones are the following:

1) Each call to sendnewportmapping is now followed by a call to
   readResponse. This is important because libnatpmp expects the
   response to each request to be read before another request is sent;
   skipping a request can cause the responses to later requests to be
   incorrect.

2) All requests are now sent from the Asio thread. This wasn't the case
   before, and as a result the responses to requests made from different
   threads would sometimes get mixed up (e.g. if thread 1 sends request
   A and thread 2 sends request B, then thread 1 could end up reading
   the response to request B).

The other (minor) issues are:

3) The readResponse function would return the NATPMP_ERR_SOCKETERROR
   error code when the MAX_READ_RETRIES limit was hit instead of
   NATPMP_TRYAGAIN, which made this case look identical to a poll error
   in the logs. The function would also sleep unnecessarily after the
   last iteration when the limit was reached.

4) The sendMappingRequest function tried to make up to MAX_READ_RETRIES
   calls to readResponse when it detected that it had gotten a response
   of the wrong type, but this doesn't work because libnatpmp doesn't
   allow reading multiple responses in a row and will simply return a
   NATPMP_ERR_NOPENDINGREQ error for each call after the first one.

GitLab: #33
Change-Id: If8f0824eef1de601411c9e70bd33df46230a24f8