[BUG] Default EPMoELayer group is rejected by get_ep_size #6485

The new EPMoELayer in #6477 accepts ep_group=None as the default WORLD group, but writes None into its expert parameters' ep_group attribute. get_ep_size(param) rejects that attribute with an assertion, even after the distributed process group has been initialized.

This affects feat/moe-ep-layer at 055853017e23c778402b2c8d965838cb09fb43e4; the new layer is not yet on main.

Reproduction in a CPU/Gloo environment:

import tempfile
from pathlib import Path

import torch.distributed as dist
from colossalai.moe import EPMoELayer, ReferenceExperts
from colossalai.tensor.moe_tensor.api import get_ep_size

with tempfile.TemporaryDirectory() as directory:
    dist.init_process_group(
        "gloo", init_method=(Path(directory) / "rendezvous").as_uri(), rank=0, world_size=1
    )
    try:
        layer = EPMoELayer(4, 8, 4, experts=ReferenceExperts(), device="cpu")
        assert get_ep_size(layer.gate_up_proj) == 1
    finally:
        dist.destroy_process_group()

Expected: the layer, its default dispatcher and its expert parameter tags consistently identify WORLD, and the getter returns 1. Actual: AssertionError: The tensor does not have expert parallel group.

The correction should preserve explicit subgroups and construction before distributed initialization. Verified environment: Ubuntu 22.04.5, Python 3.10.12, Torch 2.13.0+cu130, zero visible GPUs.