Repeating small tests many times was order of magnitude slower. The reason is the computation of
Lagrange polynomials was both inefficient and performed every time an RegularInterpLagrange instance was
created. This causes overhead which is not relevant for large problem sizes but makes running many small
tests slow. The fix for now is to cache the creation of the polynomials. Now they are created only onnce
for each interpolation order, but not once per instance.

Now a positive shift generally means a shift to the right, i.e. a delay is a negative shift.

Add relative shift location selection to interpolator interface. This will allow wrapping dsp.timeshift.

This includes a rewrite of the Lagrange interpolator using a new algorithm based on FIR filters.