refinements in preparation for LSST

[jtmyles]

This PR implements the following changes:

SOMPZ methodology

1. implement optional use of overlap_weight to assign galaxies weights in the n(z). This enables weighting galaxies by, for example, shape noise and shear response.
2. implement optional use of 'bin conditioning'. This refers to Equations 4, 5, and 6, of [Myles & Alarcon et al. 2021](https://doi.org/10.1093/mnras\
   /stab1515). The truth (equation 4) can be approximated with 'bin conditioning' (equation 5) or the more relaxed criterion of equation 6. Prior to this PR, the code only enabled equation 6. In practice, for a given dataset equation 6 may be an overly permissive approximation of the truth. The SOMPZnz class now allows the user to use the more conservative equation 5.
3. implement optional use of p(z) from calibration data rather than point estimates to learn the redshift color relation p(z|c)

both of these features were used for the DES analyses, so have been enabled in rail_sompz for future applications of rail_sompz to data and simulations.

Code architecture

4. use the parameter name nprocess for both the inform and estimate stages, instead of differing terms nproc for inform and nprocess for estimate
5. make exponentiation of Eqn A5 in Sanchez et al. 2020 overflow safe
6. implement n(z) function SOMPZnz_fast that avoids use of pandas for faster output. This should work as long as previous stages have already been ru
   n.
7. save tomographic bin assignments per-galaxy to an array (previously, only assignments per cell were saved to disk) for convenience