Inconsistent GWB from sampled populations

[lzkelley]

Same issue as noted in #21. There seems to be a systematic offset in the GWB amplitude, where the GWB from a sampled (kalepy.sample_outliers) population is ~10% higher than from the purely continuous, SAM calculation (sam.gwb()). Reason is unclear.

[lzkelley]

When not sampling any outliers, using all representative points, and also no poisson sampling, then the amplitudes match up nicely. But note the exact level of consistency depends on the number of bins. For example:
(30, 31, 32):

vs.

(150,151,152)

and note also that the overall amplitude is also changing.

With Poisson draws, the match is also pretty good, with the "sampled" version being a little bit further below the SAM/binned calculation. This is probably because weights < 1.0 values are being used when there is no Poisson draw.