Reduce flakiness of tests

[MichaelChirico]

I observed 1-2% flakiness in the test suite.

• The default nbagg = 25L produces a never-out-of-bag observation about .05% of the time, n=30 should be about an order magnitude rarer and not slow down the test much
• {ipred}'s predict.classbagg method uses sample() in the case of ties, leading to a 1-2% rate of the predictions differing purely by chance. (I suppose an alternative is to use a different aggregation=)

All told I see 1/10000 failure with this change (the 1 would go away at nbagg=35L)