Fix errors for single-node graph

[ultmaster]

This PR addresses some problems when the graph only contains one single node.

To reproduce this issue:

from torch import nn
from nn_meter import load_latency_predictor

predictor = load_latency_predictor('cortexA76cpu_tflite21')
base_model = nn.Linear(3, 3)
lat = predictor.predict(model=base_model, model_type='torch', input_shape=[1, 3])
print(lat)

Two problems are hidden inside:

1. The graph has a special initializer in this case (which also has a different dims format). I'm not sure whether it's related to the specific single-node graph, or related to my PyTorch version (v1.12.1).
2. refresh() erroneously marked the single node as unused and deleted the node, making the result always 0.

Tests should be added for this case, but unfortunately I didn't find a seat for it in the currently testing framework, so I decided to leave it for now.