Implement deconvolution as a step (and reconvolution as a layer?)

Transferring this from meeting notes. Some additional details may need to be transferred or investigated, and scope decided. E.g., convolution kernel could be (a) learned from line list or line-list-like data, with indicators and timestamps for all relevant events, (b) from separate time series, (c) set from values taken from literature or separate analysis, ...; methods can be parametric or nonparametric; etc.