Add care method

[Jamie001129]

Overview:
This is the implementation of CARE, a multi-objective, model-agnostic recourse method that generates counterfactual explanations by optimizing seven objectives using NSGA-III genetic algorithm. The method organizes objectives into four hierarchical modules that can be progressively enabled.

Paper:
CARE: coherent actionable recourse based on sound counterfactual explanations
By Peyman Rasouli, Ingrid Chieh Yu
International Journal of Data Science and Analytics (2022)

Implementation:
Dataset: the authors' version of "Adult" (32561 entries, 14 features)
Model: GradientBoostingClassifier from sklearn, which is the only model that the authors used in their experiments.
Evaluation: N=50 test instances, which already took quite a while (paper used N=500)

Key dependencies:
numpy==1.21.6 # (upgraded from 1.19.4 for hdbscan compatibility)
deap==1.3.1 # NSGA-III optimization
dython==0.6.1 # Coherency objective
cython==0.29.37 # Build dependency for hdbscan
hdbscan==0.8.33 # Connectedness objective
NOTE: I am yet to modify the requirements-dev.py in my branch

Table 7 reproduction results:

Config	Source	↓O_outcome	↓O_distance	↓O_sparsity	↑O_proximity	↑O_connect	↓O_cohere	↓O_action
{1}	Original	0.00±0.0	0.02±0.0	1.41±0.7	0.58±0.5	0.20±0.4	0.05±0.2	0.08±0.3
	Reproduced	0.00±0.0	0.03±0.0	1.46±0.8	0.66±0.5	0.18±0.4	0.13±0.3	0.12±0.4
{1,2}	Original	0.00±0.0	0.12±0.1	2.84±1.7	1.00±0.0	1.00±0.1	0.10±0.3	0.36±0.6
	Reproduced	0.00±0.0	0.13±0.1	3.00±1.7	1.00±0.0	1.00±0.0	0.13±0.3	0.34±0.6
{1,2,3}	Original	0.00±0.0	0.12±0.1	3.05±1.7	1.00±0.0	1.00±0.1	0.00±0.0	0.35±0.6
	Reproduced	0.00±0.0	0.13±0.1	3.20±1.7	1.00±0.0	1.00±0.0	0.00±0.0	0.44±0.7
{1,2,3,4}	Original	0.00±0.0	0.11±0.1	2.76±1.8	1.00±0.0	0.91±0.3	0.00±0.0	0.00±0.0
	Reproduced	0.00±0.0	0.12±0.1	2.96±1.7	1.00±0.0	0.90±0.3	0.00±0.0	0.00±0.0

I expect better reproduction results if run on more instances.