Deglitch read chains

This ended up being an action-packed branch. The goal was to change BayesWaveGlitchClean to read chain samples and compute TD wavelets directly, instead of reading the FD wavelets from file, and upsampling/zero-padding/iFFT'ing the glitch model for subtraction from the frame.

Along the way we realized that there was an inconsistency in the complex phase conventions between the code and the BW paper and intended wavelet model, which are now fixed.

While fixing that, we got wild and replaced the exp() calls in the wavelet amplitude (which were dominating the run-time for the signal/glitch part of the model) with recursive relations. Now BayesLine is in the critical path for improving the cost/cycle.

The good news is that (a) the code is faster and (b) running the cleaning step is now easier to integrate into the pipeline.