refactor: Move Peaks from AOS to SOA

[jspaezp]

#166 (comment)

This PR removes the MS1Spectra and the generic peak/imspeak in ProcessedSpectrum

LMK if this is what you had in mind ... RN I am not the biggest fan of a couple of places in the processing where I am zipping->unzipping to sort the peaks (which is why I am leaving it as a draft for now).

In my benchmarking it is almost as fast as the other implementation (runtime is maybe 2% slower in my computer and that matches the number of instructions retired on some random data)

LMK what you think! (or feel free to take it as a base to go off ...)

[lazear]

Thanks for the draft - while it might not be faster yet, it should unlock some better performance both in theory (cache locality, esp for binary search) and enable some further downstream optimizations.

Some ideas (not necessarily for us to implement now):

• Reading from columnar data (naturally SOA), for example mzparquet
• SOA could enable something like inline compression of data to further reduce RAM consumption
• There are some optimizations around binary search (e.g. select_most_intense_peak) that could leverage the fact that we know the mz array is sorted

[jspaezp]

LMK if you want me to polish it to be a "real PR" (I can also work with the authors to rebase #173 and #181 ... since I would be adding again merge conflicts to those ... I think ...)