Refactor matrix inversion

[brandynlucca]

This PR includes updated functions, methods, etc. for running the matrix inversion analysis (for both krill and in general).

Note: the tests have pretty spotty coverage at the moment so I am still actively working on these to make sure at least the primary usage is covered and working from there.

Name                                                       Stmts   Miss  Cover
------------------------------------------------------------------------------
echopop\inversion\inversion_matrix.py                        192    117    39%
echopop\inversion\operations.py                               65     35    46%
echopop\inversion\scattering_models.py                        95     82    14%

[brandynlucca]

Did you observe significant benefits to go with JIT compilation for the calculations under scattering_models.py?

Yep! It's possible that it was just my machine, but I was running into significant performance issues when running the optimization step via lmfit. So when I profiled the code, the pcdwba and pcdwba_fbs functions expectedly overshadowed the other operations as being the costliest. For the overall pcdwba function, a single run of the function was taking upwards of ~2 seconds per call, which when repeated over the optimization results meant a single transect (i.e. a single parameter set) was sometimes taking up to ~16 minutes (extrapolated out to ~27 hours per 100 transects).

When profiling those functions, the Bessel function interpolation was heads-and-shoulders above everything else. Adding the @njit(fastmath=True) decorator around the wrapper function batch_interp conferred a 10-fold decrease in the calculation time for pcdwba_fbs. For whatever reason, the phase calculation was also a major contributor to the bottleneck despite being a pretty straightforward calculation. Wrapping that calculation with the fast_phase function with @njit decorator effectively negated how much the calculation contributed to the slowdown.

From what I understand, I think converting the frequency-loop within pcdwba_fbs to be a JIT-compiled function would probably have been the better approach, but I didn't want to overcomplicate things for the time being until there is a clear need for improving performance (i.e. not now). I just made sure that I could at least make it through a full set of model runs on my end of things so I could actually bugtest and identify edge cases that I hadn't anticipated.