[BC-Breaking] TuSimple max lane number testing constrain bug fix

[voldemortX]

This bug fix brings 0.1~0.4% F1 score increase on all uploaded models (on one extreme case ERFNet Baseline, 0.8%).
The original code did not mask out the lane prediction with lowest existence probability, but it accidentally masked out the first lane when the network predicted 6 lanes (TuSimple has a maximal number of 5 lanes).

This change is only applied to testing scripts, so any previous downloaded/trained weights still can be used, they just need re-testing. Also this change is too small to have much of a effect on hyper-parameter validation, so hyper-parameters will remain unchanged as well.

However, we are providing new best model weights for download (since the previous ones now are not the best among the 3 runs after re-testing), these models are:

erfnet_baseline_tusimple_20210201.pt
resnet101_baseline_tusimple_20210217.pt
resnet34_baseline_tusimple_20210216.pt
resnet18_scnn_tusimple_20210216.pt
resnet34_scnn_tusimple_20210216.pt

Be sure to re-download the new best models (filename ends with _20210424.pt).