Fix of empty leaves issue in MT building and adding a flag to force MT building on one GPU even in multi-GPU systems

[dloghin]

1. In new_from_leaves_2d(), we performed a flattening of the 2D Vec<Vec> into 1D Vec. However, if some inner Vec is empty, the final number of leaves in the 1D array is wrong. We fixed this by adding F::ZERO when the leaf is empty.
2. We add a FORCE_SINGLE_GPU flag to force building the MT on a single GPU even if NUM_OF_GPUS is > 1.