fix: convert grouped-expert fc1 bias together with its FP8 scale

[shiaho777]

Loading linear_fc1 for local experts aborted when add_bias_linear and a block scale were both set. The weight path already reads gate_up_proj_scale_inv, and the bias path already copies gate_up_proj_bias with hf_scale_inv left empty. The assert was the only thing keeping the two apart. Export already writes the bias and the scale as separate tensors.

The assert now only requires a local expert. A non-expert fc1 bias is still refused, because that layout is not converted here.

flake8 is clean on gpt_bridge.py.

[hjh0119]

Which model actually encountered this issue?