s3: Fix a segfault in the RPC server

s3: Fix a segfault in the RPC server

After converting the rpc infratructure to talloc, read_from_internal_pipe freed
the outdata too early. If the last fragment was read in two pieces (as
rpcclient does it), all the outdata was freed during the read of the first
piece of the read of the last fragment. Later read&x calls, trying to read the
rest of the last fragment stepped into p->out_data.frag with non-zero offset
when this was already freed.