Draft: Snr chisq hist

Overview

1. Move train_noise to StrikeSink
2. Construct SNR-chisq tensors on the GPU, only transfer the histogram to the cpu once at the end of the pipeline
3. Performance (GPU):
   1. without old train nosie: ~300,000 templates per core
   2. with old train noise: ~200,000 templates per core
   3. with this MR: ~300,000 templates per core

Todo

1. Sanity check between old and new results.
2. I am not sure if I understood what count tracker is trying to do, and it is slow if there are triggers above the count tracker criteria.