reorganize code: PAL + SAR as one integrated repository

0.01 magnitude precision

Add location module

1. better default values, especially for glitch removal
2. adaptive neg_ratio (one less param in config)

1. Add self-attention after RNN, thus the model is now named as "Self-Attention RNN (SAR)"
2. Modify neg-cut strategy, high-assoc sta-date pairs are weighted down, so that is more suitable for intense sequences
3. Add glitch removal algorithm to realize higher stability
4. More suitable default parameters, should have ~2.75*PAL detection number

Use RNN for Sequence Labelling (RSeL). The RSeL model is trained under TED workflow, with PAL detections. RSeL output P&S probability, as most AI pickers do. The RSeL picks can be associated and located with PAL toolkit

Bug fixed

1. New name: Training-base Earthquake Detection (TED) workflow;
2. New CNN model: ResNet style;
3. New training sampling method: use unassociated PAL picks as Glitch (the second kind of noise), which is suitable for interseismic period

Better usage; bug fixed

Major updates to realize an easy usage