WIP: better Meyer FIR approximation

[grlee77]

Do not merge this PR. It is just a demo of generating more accurate FIR approximations for the Meyer wavelet (see #261). We can choose one of these to replace the current approximation.

The 102 coefficient approximation here is in agreement up to 4 or 5 digits with the filter coefficients used by Matlab. The 66 coefficient approximation is as in the reference provided by @holgern. And is close in accuracy to the 102 coefficient case. Note that the sample points are shifted by 0.5 between these two cases.

Both of these have much better accuracy than the current approximation:

Error (pywt dmey) = 0.0036067702113178066
Error (dmey102) = 4.351866673799496e-06
Error (dmey66_async) = 9.899204593989887e-06

If the asynchronous sampling case is extended out to 102 samples, accuracy is further improved.

Error (dmey102_async) = 9.934928444371677e-07

The 66 sample case is probably the best computation time / accuracy tradeoff.

[rgommers]

as just discussed: likely makes sense, but let's not look at this till after 1.0