Revise _negbinom_ estimation and reorganize `make_forecast`

Revise negbinom estimation and reorganize make_forecast

This PR is two-fold, with both aspects depending on each other somehow.

---

Details:

1. Change negbinom implementation

ATTENTION: this affects case H of floatCSEP!

• correctly determine counts
  • by using the full catalog span from the date of 1st event (catalog_start.date(), any magnitude) until 'now' (t0)
  • to indeed end at 'now', need to adjust the last bin edge created with np.arange()
• correctly estimate tau parameter (it's 1 / alpha, not 1. / alpha * mu)
• correctly determine mean of the background/'past' seismicity (mean_bg)
  • it's not the mu or lambd estimates for Poisson, since negbinom sampling must refer to the mean 'n_events' above mag_min, not above mag_compl adjusted to mag_min
  • essentially, mean_bg is just the average of the counts array
  • So why can mag_compl not be used in negbinom parameter estimation? Because it is based on a variance estimate (of events >= mag_min), which cannot be adjusted to a different magnitude threshold. I consider this is a weakness compared to using Poisson; perhaps there is a clever solution?
• calculate parameters for 'recent' seismicity component differently than for background/'past' seismicity
  • since variance for 'recent' seismicity is biased or not defined, use background's dispersion to estimate it from the 'recent' mean
• outsource negbinom parameter estimation to a function to facilitate treating background and recent seismicity differently (only based on mean and var)
• further simplifications
• add some assertions and detailed comments (purely informative)
• FYI: former condition if lambd != 0: ... else ... (no recent seismicity) is now caught by:
  1. if mean == 0 : return 1, 1 (tau, theta) in calc_negbinom_params(),
  2. leading n_events = numpy.random.negative_binomial(tau, theta) to always sample a zero.

---

2. Reorganize make_forecast, incl. completely separate Poisson/negbinom and use clear names for catalog subsets

Poisson-based and negbinom-based modeling share some catalog subsets for sampling event locations, but use different magnitude-thresholded catalog subsets for parameter estimation (see above). Clarifying those relations necessitated some reorganizations; specific changes:

• create and use catalog subsets with self-explanatory names:
  • cat_past (past events)
    • → cat_past_compl (formerly cat_total, above mag_compl), used by Poisson and always for sampling event locations
    • → cat_past_minmag (formerly cat_total_mag, above mag_min), used by negbinom
  • cat_prev (recent events)
    • → cat_prev_thresh (formerly catalog_prev, above mag_compl or mag_min depending on apply_mc_to_lookback), used by Poisson and always for catalog sampling
    • → cat_prev_mingmag (recent events above mag_min), used by negbinom
• completely separate Poisson- and negbinom-related operations (previously, the latter still depended on the former, see above)
• rename extension apply_mc_to_lambda to a more general apply_mc_to_lookback ('lambda' suggests compatibility only with Poisson, but sampling locations of recent events above mag_compl can also be used with negbinom).

  • ATTENTION: this also affects case H of floatCSEP! Please adapt models.yml!

• More parameter explanations in args.txt
  • incl. that mag_compl and apply_mc_to_lookback input arguments are not used in negbinom parameter estimation.