Add TVComplement metric

[katxiao]

Problem Description

Add a TVComplement metric. For each discrete column: Calculate the frequencies of each category in the real data (r0, r1, …), expressed as a proportion in [0, 1]; also calculate frequencies in synthetic data (s0, s1, …). Calculate the Total Variation Distance between the two distributions and then normalize it.

The breakdown only contains the score.

Expected behavior

# single table case
>>> TVComplement.compute_breakdown(real, synthetic, metadata)
{
  'user_age': { 'score': 0.96667 },
  ...
}