Despite similar timing, this is different from #4111 by only affecting ip6 tests but across multiple platforms. Contrary to #4114 this regression is generally less than 5%. The best signal is in this [0] trending, but variance of trial within run is way less than variance of averages between runs. First few bisects were too noisy, not worth investigating further.

[0] vpp-2n-spr-200ge2p1cx7veat-mlx5-ip6-78b-1c-ethip6-ip6base-mrr