fix InputFeatures .to

[NicholasCao]

In the following two scenarios,

for step, inputs in enumerate(cycle(train_dataloader)):
    inputs = inputs.to(device)
    ...

or

dataloader = cycle(iter(train_dataloader))
for step in range(max_steps):
    input = next(dataloader)
    inputs = inputs.to(device)
    ...

raw InputFeatures.to(...) causes infinite growth of GPU memory and unavailability of used inputs (e.g. SoftTemplate modifies input_ids of inputs by batch['input_ids'] = None)