Optimize query_non_zero_columns()

[cberner]

This reduces the time spent in the fourth phase from ~6% of encoding time to ~1%, according to perf, and improves overall throughput by 3-4% on large symbol counts.