WIP: Improve likelihood informed threshold value

[odunbar]

Purpose

The theory indicates that actually one should be truncating likelihood information based on the retained spectral mass, This tends to give better cut-offs that the current approach.

To-do

• The output-space threshold should be updated to something amenable to greedy algorithms

Content

• We also rename the kwarg to retain_spectral_mass from retain_info
• Improve the diagnostics upon truncation
• Modify the cut-off to produce the smallest k such that
  $$\frac{\sum _{i\leq k} \log(1+\lambda_i) }{\sum _{i} \log(1+\lambda_i) } \geq T $$
  This basically removes the square on the summand of what is currently done