fix when only one ped in scene

[test-bai-cpu]

Now in trajnet++, all scenes with only the primary pedestrian in the scene, will be removed. When I add these scenes, in the training step, it will have an error:

return torch._C._nn.linear(input, weight, bias)
RuntimeError: mat1 and mat2 shapes cannot be multiplied (8x36 and 288x256)

Because in trajnetbaselines/lstm/gridbased_pooling.py, grid, generated by occupancy() , the shape of grid maybe wrong in single pedestrain's case. The grid shape is [num_tracks * batch_size, self.pooling_dim, self.n, self.n], like [32,2,12,12] but in code for only one pedestrian, the grid shape is [num_tracks, self.pooling_dim, self.n, self.n], like [1,2,12,12]. It should be [8,12,12,12] in this case.