Needed to filter `nans` in `solve_mab`

[motiwari]

Otherwise, if there's a single nan, all arms are filtered because nothing is less than np.nan.

This only happens in classification.

An alternative error is the $\epsilon$-suboptimality we allow

[motiwari]

I've pushed code to reproduce this in the branch repro_random_sample_bug.

To reproduce the results, run python compare_runtimes.py from FastForest/experiments/runtime_exps and see the output:

 Running runtime comparison for: HRFC
Fitting tree 0
Number of candidates survived:  0 ucbs.min():  nan
Terminated after  1  rounds
Calculated split with 8000 queries
Fitting finished
Fitting tree 1
Number of candidates survived:  0 ucbs.min():  nan
Terminated after  1  rounds
Calculated split with 8000 queries
Fitting finished
Fitting tree 2
Number of candidates survived:  0 ucbs.min():  nan
Terminated after  1  rounds
Calculated split with 8000 queries
Fitting finished
Fitting tree 3
Number of candidates survived:  0 ucbs.min():  nan
Terminated after  1  rounds
Calculated split with 8000 queries
Fitting finished
Fitting tree 4
Number of candidates survived:  0 ucbs.min():  nan
Terminated after  1  rounds
Calculated split with 8000 queries
Fitting finished
Ours fitted 1.2937331199645996
Fitting tree 0

A fix for this already exists in the branch random_baseline, which I'll merge into dataset_experiments

@vxbrandon