fix(server): loopback-bind the TP rendezvous store, add --dist-port

[shikhakath]

Fixes #301.

The TP rendezvous port was hardcoded to server_port + 1, so two instances on adjacent ports fight over it. There's also a second issue the thread surfaced: even though the init URL says tcp://127.0.0.1:PORT, torch's TCPStore ignores the host you give it and binds every interface -- so today it's an unauthenticated store reachable off-box on any bare-metal or --network host setup.

This keeps TCP (per the thread's discussion, since a future multi-node rendezvous still needs it) but builds the store explicitly instead of trusting
init_method: rank 0 pre-binds a socket to 127.0.0.1 and hands the fd to
TCPStore via master_listen_fd. Other ranks connect as normal clients.
--dist-port makes the port configurable, defaulting to --port + 1 so nothing changes for existing users.

Testing

I don't have a GPU, and couldn't even install the package to run it locally (flashlib has no macOS wheel) -- so no checkpoint, no ft serve, no driver to report.

What I could check: I reproduced the exact logic standalone -- pre-bind a loopback socket, pass it to TCPStore via master_listen_fd, spawn two processes and run init_process_group + all_reduce over gloo. The socket only binds to 127.0.0.1, and the rendezvous still works. I've added that as two real tests
(tests/engine/test_distributed_store.py, tests/server/test_dist_port_arg.py),
but haven't run them against this actual package -- someone with an install should confirm they pass before merge.

Completely untested: the NCCL/TP>=2 path in _init_communication. Same gap nobody in the original thread could close either. Happy to test on a multi-GPU box if someone points me at one.