1. For memory usage, deprecate Docker(runc) container runtime.
Instead use our own native implementation, in C module "pyctr",
containing single source file pyctr.c, only support init and exec.

Docker's implementation require a shim for each container,
which consumes considerable program memory in user space.

Besides, docker need complex ipc chain like
docker<->dockerd<->containerd<->shim<->runc, slowing container init.

Even some trials on lower level api (runc/crun),
there's still unaffordable memory consumption in kernel.

Finally, I reinventing the container management wheel specific to our
environment, comprising some native linux syscall, just for less memory
and a little improvement in container initialization.

2. For network, use veth in place of Docker's veth-bridge-veth mode.
Docker's default network scheme occupy more CPU on softirq
and we don't need it in satellite point-to-point link.

NEED SOME EXTRA WORK
change:
1. refactor unnecessary Observer class since we only use it once.
Functions with no state may be better.
2. change link representation from matrix to human-readable list.
Such sparse structure can also save much memory and disk.
3. use more numpy ndarray operation for better performance(maybe).
4. abstract interfaces for more flexible link style
5. use individual bird.conf instead

NEED SOME EXTRA WORK

distance, check route, iperf, sr, damage & recovery
update node representation to support GS operation

fix: typos

Support for arbitrary topology