Allow flattening in Kullback-Liebler distance

[siddhartha-gadgil]

• Otherwise too little weight is given to even types with moderately small probability.
• Here flattening means taking a power (less than 1).
• An alternative is to have coefficients higher in synthetic generation - this is not done in actual generation because of risk of blow up. This is not equivalent to the above, depth is still penalized.