fix: clone tensors in recv_state_dict to avoid aliasing

[Abhishek21g]

Fixes #245

Problem

recv_state_dict receives tensor payloads into buffers derived from og_state_dict, then returns a state dict whose tensor values alias the caller's template objects (id match).

This affects live checkpoint recovery in checkpoint.py (recv_ckpt_from_peer), where optimizer/training state is received via recv_state_dict(..., self.optimizer.state_dict()).

Repro

Identity check from the issue thread — after receive, returned tensors share objects with og_state_dict:

received = recv_state_dict(pg, src_rank, og_state_dict)
assert id(received["optim_sates"]) != id(og_state_dict["optim_sates"])  # fails on main

Fix

Clone received CPU buffers into a fresh tensor list before _load_sendable_state_dict, so the returned dict owns independent tensors.

Also:

• add test_recv_state_dict_no_tensor_aliasing (no distributed required)
• fix send_state_dict / recv_state_dict call signatures in the skipped dist test
• fix dist test to assert received values match the sender payload

Verification

uv run pytest tests/test_dist/test_send_state_dict.py::test_recv_state_dict_no_tensor_aliasing -v
uv run pytest tests/test_dist/test_send_state_dict.py::test_load_state_dict -v

Dist test (test_send_recv_state_dict) remains skipped (hang); happy to help unskip on a 2-process Linux host separately.