Bug modification about dense_image_warp function

[luochuyao]

the function detail of gather() is different with tensorflow when it implements interpolate_bilinear()

I use this :

linear_coordinates = torch.reshape(linear_coordinates,[batch_size*num_queries])
gathered_values = torch.index_select(flattened_grid, 0, linear_coordinates)

to replace

gathered_values = torch.gather(flattened_grid.t(), 1, linear_coordinates)

at 342 line

It should be note that