Reduce flakiness of test for Eigen::half.

Some tests got really flaky for Eigen::half with recent updates to the implementation for AVX512FP16.

In particular,

  • the produce_large test when comparing a vector dot products to a v1.cwiseProduct(v2).sum()
    • the true product uses FMA while the latter doesn't, involving lots of intermediate rounding for Eigen::half
    • resolved by forcing scalar FMA in the comparison
  • the packetmath test reduce_mul resulted in very small numbers since we were multiplying many values < 1
    • adjusted the range of values to reduce flakiness
  • AVX512 without FP16 was missing pmadd functions for Packet16h resulting in a failing test

Merge request reports

Loading