Primarily this fixes an off-by-one buffer overwrite (rare but still existing).

I also switched from calloc() to malloc() as a minor performance boost since
the rest of the code fills in the structs fine anyway - and they must for the
case when we use the stack-based auto variable array instead of the allocated
one.

I made the loop filling in poll_fds[] break when poll_nfds is reached as a
minor speed improvement.