fix shape bug in lfbo pattern search

[ethche]

After running some additional benchmarks, I realized that for some kernels, there are many configs that give identical performance (e.g. for rms_norm). Originally, LFBO Pattern Search construct training labels by classifying which configs have performance strictly better than quantile(perfs, 0.3). Since this comparison is strict, in cases where the performance of the top 30% of configs is identical, then all the labels are zero. Training with these labels doesn't throw an error but instead when we call predict_proba, this returns all a 1D tensor of all zeros instead of a 2d tensor of probabilities for each class.

Now this is fixed by constructing labels with a non-strict comparison. We also explicitly deal with the case where all the labels are identical by arbitrarily flipping the first one.