far.py: don't convolve snr,chi numerators with KDE

- (snr, \chi^2) numerators are computed analytically.  previously, changes
  to the KDE used to estimate the denominator PDF from sample histograms
  could change sensitivity measurements simply because of the effects those
  changes were having on the numerator signal model.  this patch removes
  that confusion.
- because the limited set of discrete prefactors used for constructing the
  numerator creates a stair-step effect, some amount of smoothing is
  desired in the numerator regardless.  this is smoothing, not density
  estimation, they are distinct from one another although the same code has
  been used to accomplish both in the past.  this patch moves that
  smoothing to where the source of the banding is in the code to
  disentangle it from the background density estimation.  if the prefactor
  banding is adjusted, that smoothing can be adjusted as well without
  entangling it with the denominator's kernel design.
- refs #2339