Release fds and the multicast handle on Buffer construction error paths

[yurekami]

Fixes #17.

Guards the acquire-without-finally windows reported there, on all four sites, using only bindings the file already imports:

• _exchange_ipc_fds: if the recv loop raises (timeout, or the explicit no-fd RuntimeError), fds already duplicated into this process by SCM_RIGHTS are now closed before the exception propagates. sock.close() in the existing finally never covered them; they are plain ints with no Python owner.
• _map_nvl_dist_tensor: os.close(local_fd) moved into a finally around the exchange, so the exported fd cannot leak when the exchange raises.
• _create_nvl_multicast_view: the window between nvl_multicast_create and a completed import is now wrapped; on error the root closes its exported fd (finally) and releases the multicast handle via the existing nvl_release_mem_handle binding (a multicast object handle is a CUmemGenericAllocationHandle, so cuMemRelease applies). Success-path behavior and timing are unchanged.
• create_nvl_single_owner_tensor: same finally treatment for the owner's exported fd.

A leaked exported fd also pins the underlying cuMemCreate allocation (release only completes when all shareable references are gone), so on the production path this held the hidden_buf chunk, not just an fd table entry — details in #17.

Verification: behavior verified by reading against master @ 2bd860b plus py_compile; I don't have a multi-GPU NVLink box available, so the error paths were not executed here. Happy to adjust if you'd rather scope the multicast guard differently.