Fix a trio of bugs in the local/remote window calculation for datagram

Fix a trio of bugs in the local/remote window calculation for datagram
data channels (i.e. TunnelForward):

Calculate local_consumed correctly in channel_handle_wfd() by measuring
the delta to buffer_len(c->output) from when we start to when we finish.
The proximal problem here is that the output_filter we use in portable
modified the length of the dequeued datagram (to futz with the headers
for !OpenBSD).

In channel_output_poll(), don't enqueue datagrams that won't fit in the
peer's advertised packet size (highly unlikely to ever occur) or which
won't fit in the peer's remaining window (more likely).

In channel_input_data(), account for the 4-byte string header in
datagram packets that we accept from the peer and enqueue in c->output.

report, analysis and testing 2/3 cases from wierbows AT us.ibm.com;
"looks good" markus@