Debug: 1. cache refresh at every init; 2. compute cardinality only once

[MarkDana]

Bugs solved:

1. About cache refresh:

• Manually refresh cache at every init. Added one refresh line at SkeletonDiscovery.py:L145 (for PC) and FCI.py:L615 (for FCI).
• Deprecate hash(data.tobytes()) - it's slow. @chenweiDelight also mentions a faster hash(str(data)) - but let's just init refresh and use no hash.

2. About passing cardinality chisq or gsq:

• Added cardinalities and is_discrete at Fas.py:L10, and calculate cardinalities only once at SkeletonDiscovery.py:L165 (for PC) and FCI.py:L625 (for FCI). No need to np.max() every time.

TODO: still debugging:

Now PC results with/without fas (change PC.py:L71) respectively:

data (#nodes/#edges) time (sec)	without fas time	with fas time	SHD(without fas, with fas)
cancer 5/4	0.009	0.009	0
earthquake 5/4	0.012	0.01	0
survey 6/6	0.014	0.015	0
asia 8/8	0.024	0.025	0
sachs 11/17	0.142	0.148	0
child 20/25	0.618	0.699	4
insurance 27/52	1.45	1.823	3
water 32/66	0.321	0.365	0
alarm 37/46	0.873	1.015	1
barley 48/84	3.466	4.95	0
hailfinder 56/66	0.938	1.588	0
hepar2 70/123	9.482	11.949	3
win95pts 76/112	3.381	4.504	0
andes 223/338	26.741	45.722	0

So two problems:

1. Still small SHD difference at datasets e.g. child.
2. Still time difference, e.g. andes 26s vs 45s - though before solving cardinalities issue, it's ~300s.
3. See profiling stat: callings count on chisq is different on with/without fas.

[MarkDana]

Closed this pr - the two issues (cache for different data & pass cardinality to chisq/gsq) has already been covered in @chenweiDelight 's newest commit (732b9d1).