New thresholds in table

[Torbjorn Granlund]

I am not sure HGCD_APPR_THRESHOLD tuning works reliably yet.

If alg A wins over alg B for sizes < m and for sizes >= n, then the only
reliable way of determining n seem to be to first determine m.

Measuring from above might work, if one can guesstimate the highest
reasonable value, but this might be quite slow.

Then again, if alg A is faster than alg B below n, perhaps we should use
alg A in for such small sizes?

-- 
Torbjörn