Numba JIT classes

[rfayat]

Hi !

First of all thank A LOT for this great collections of algorithms and the super neat API and documentation that come with it. It has proven very useful to me to test a bunch of AHRS filters on my data.

I've been experimenting a bit with numba's JIT classes, which allow to compile parts of the code with only a few changes, and it looks like it can greatly improve computation time for some algorithms (E.G. -90% for EKF on a MARG time series with 100k samples).

@Mayitzin made it pretty clear that the goal of the package is not efficiency but since I needed to run the pipelines on several files and the computation time got a bit long, I converted a few algorithms (Mahony, Madgwick, EKF) to JIT classes in my fork of the repository. A small benchmark of the computation times is available below for time series with 100k samples for IMU data or also with 3-axis magnetometers (MIMU):

	Pure Python (IMU / MIMU)	NumbaClass (IMU / MIMU)	Δt (IMU / MIMU)	Overhead
Madgwick	8.93s / 15.31s	3.33s / 4.39s	-62% / -71%	9.17s
Mahony	14.91s / 24.08s	5.38s / 4.90s	-64% / -80%	21.34s
EKF	73.26s / 134.95s	11.91s / 11.72s	-83% / -91%	32.33s

I don't intend to open a PR since this conversion to jit classes doesn't really fit the spirit of the package (it adds an overhead due to the compilation and also a dependency) but I just wanted to share this here in case some people have similar needs as me and want to try out the numba-compiled version of the code.

Once again thanks a lot for the great work !

Cheers

rfayat