Update feature imputation strategy

[bfhealy]

Missing features are current zero-imputed for training and mean-imputed for inference. Aside from their inconsistency, there are better (albeit more costly) ways to perform imputation. The current plan is to:

• Exclude the AllWISE W3 and W4 magnitude errors, which are missing from >75% of the training sample
• Impute a value of zero for missing mean_ztf_alert_braai
• Impute the median for missing magnitude errors (mainly PS1)
• Use regression (e.g. KNN imputation) to impute missing magnitudes
• Use regression (potentially on a class-by-class basis) to impute missing Gaia EDR3 parallaxes

[bfhealy]

Using sklearn's KNNImputer (n=5) takes ~2-3 minutes to run on 100,000 sources for the following feature subset:

['Gaia_EDR3__phot_bp_mean_mag',
 'Gaia_EDR3__phot_rp_mean_mag',
 'Gaia_EDR3__parallax',
 'PS1_DR1__gMeanPSFMag',
 'PS1_DR1__rMeanPSFMag',
 'PS1_DR1__iMeanPSFMag',
 'PS1_DR1__zMeanPSFMag',
 'PS1_DR1__yMeanPSFMag',
 'AllWISE__w1mpro',
 'AllWISE__w2mpro',
 'AllWISE__w3mpro',
 'AllWISE__w4mpro']

This takes significantly longer than inference (~10s) for the same number of sources, but it remains important to change our imputation strategy from its current state.