Fix multi inputs bug.

[GeoffreyChen777]

Here is the demo to reproduce the bug:

import torch
import torch.nn as nn
import torchvision
import torch2trt

max_batchsize = 1
max_trt_workspace = 1 << 30 # 1GB

class Net(nn.Module):
    def __init__(self):
        super(Net, self).__init__()
        self.conv = nn.Conv2d(3, 64, 3, 1, 1)
    def forward(self, x, y):
        x = self.conv(x)
        y = self.conv(y)
        return x + y

net = Net().cuda().eval()

inputs1 = torch.rand(1, 3, 4, 4).float().cuda()
inputs2 = torch.rand(1, 3, 4, 4).float().cuda()
net_trt = torch2trt.TensorRTModuleWrapper(net, max_batchsize, max_trt_workspace, param_exclude=".*AuxLogits.*").cuda().eval()
out_ref = net(inputs1, inputs2)
out = net_trt(inputs1, inputs2)
print(out_ref.size())
print(out.size())
print("Diff: {}".format((out_ref-out).abs().sum()))

Reason: https://github.com/traveller59/torch2trt/blob/75f07a7d0e061ac158c937142da40dfae764d296/torch2trt/module.py#L34