instead of passing multiple tag:val pairs for single tag, provide comma-separated list: tag:v1,v2,v3 (still some tag sets to switch over)

for exec time feature, averaging 5 sets of trials and taking the minimum over those 5 averages; also warn when average and min vary significantly

larger prob sizes for dg diff mat fetch (transposed elem) variant mknl on Radeon R9 Fury to get more consistent feature costs