fix invalid ending point bug

[larksq]

1. fix a bug causing loss to generate nan and explode the training! due to wrong invalid goal point checking
2. update end point invalid checking during dataset generation
3. add multi-processing into data generation