Bug in _compute_backward_tracks

[RafailFridman]

Hi!
Thanks for your work.
I found a bug in the _compute_backward_tracks function:

There, in line 275, you write:

inv_tracks, __, inv_visibilities = self.model(
            rgbds=inv_video, queries=queries, iters=6
        )

which uses original (not flipped) queries.

I think that instead, it should be

inv_tracks, __, inv_visibilities = self.model(
            rgbds=inv_video, queries=inv_queries, iters=6
        )

Notice that I changed queries to inv_queries.

This bug is possibly related to #41 .
Also, this model call should probably accept the wind_S to make it consistent with the forward tracking model call.

I did not create a pull request, but I can make one if you accept them from other people.

[RafailFridman]

A related question:
You add a support points grid at frame 0.
But when you want to track backward, you flip all the queries, including the support grid queries, which now "start" at the last frame of the video, and are not supporting the tracking process.
Am I missing something here?

Shouldn't the support grid be recreated again to start from frame zero also in the flipped video?