Refactor CFs/Metrics to utilize numpy arrays

[samland1116]

• Optimize query performance and calculations by replacing memory intensive pandas operations with numpy-based alternatives where applicable, similar to the recent 'exceedance_probability' TCF solution

[samland1116]

Initial findings are that the difference in compute is negligible, though testing at larger scales could warrant investigation.

@samlamont any thoughts? Maybe we just want to do targeted refactoring where we have resource intensive sorts/joins/etc.?

---

small hourly data set for 2 gages, 1 config (2000-09-30 to 2012-09-29) [200,350 rows] :

• initial relative bias run time >>> 16.5s
  • updated relative bias run time >>> 18.0s
• MARE run time >>> 0.8s
  • updated MARE run time >>> 0.7s

---

large daily data set for 30 gages, 4 configs (1992-09-30 to 2010-09-28) [678,300 rows]:

• initial relbias run time >>> 26.5s
  • updated relbias run time >>> 25.8s
• MARE run time >>> 2.6s
  • updated MARE run time >>> 2.7s
• NSE run time >>> 2.8s
  • updated NSE >>> 2.7s
• KGE run time >>> 2.8s
  • updated KGE >>> 2.5s