with PAX support longer ones are allowed

fix bug concerning retval of readv (was exactly the wrong condition
check)

using 512MB for the stream buffer and copy_block_size uses up to 1.5GB
of RAM

even though I am using strncpy exactly as designed, sufficiently many
people seem to not do so and gcc complains, so I used memcpy instead

not very efficient or pretty (too many malloc/free) but enough to test
for now.

Possible improvements:
* use less malloc/free
* use ring buffer (or bipbuffer)
* use POSIX asyncio and get rid of thread (combine with eihter one of
the above)

they do not have to be the same:
https://www.mcs.anl.gov/research/projects/mpi/mpi-standard/mpi-report-2.0/node47.htm

this is a bit clearer than hard-coding the relation between the request
number and the worker it was sent to