Was accidentally using `m_max` in place of `M_max` in one spot,
thereby causing bugs.

Added support upper/lower cutoffs on component masses, as well as an
upper cutoff on the total mass.  While this shouldn't be necessary
when producing synthetic PE results, it _is_ however useful for
testing purposes, especially in the case of P-P plots, as we need to
draw the true values from the prior, and that prior must both be
proper (uniform component masses out to infinity isn't proper) and not
include points we have a hard time simulating (very low mass sources
are tougher to generate waveforms for due to duration), hence the
limits on component masses.  The limit on total masses was just added
for good measure.

These limits are all off by default and will not change results that
don't use them.

We were shifting the Gaussian incorrectly, biasing all our results. No
priors should be involved when moving the Gaussian's center, and on
top of that, we should allow for invalid masses and spins to be used
as the mean value -- we use a truncated Gaussian after that point, so
letting the center go outside the support is perfectly valid, and
disallowing it was biasing us, especially near equal mass ratio where
we had a 50% chance of going outside the range -- which was being
suppressed.

Also deleted unneeded argument

Used to be needed. Isn't with new approach.

Instead of using an ad-hoc method, where we sample
in (mc,eta,chi_eff), and then take a gigantic sample of chi12xyz
values and associate the closest match with each chi_eff sample, we
use MCMC, sampling in mc, eta, chi12xyz, and use a likelihood that
only contains chi_eff information (from the Fisher matrix).

Also deleting print debugs