Fix EM mixture-frequency optimization

[thomaskf]

Two fixes in ModelMixture::optimizeWithEM (used by -mfopt -optalg_qmix EM). The issues were reported by Ryo.

1. EM made the result worse instead of better (the final log-likelihood was lower than the start). The E-step used old partial-likelihood values that were not updated for the current model, so the per-class weights were wrong. Fix: call clearAllPartialLH() before each computePatternLhCat that builds the weights, so the values are recomputed.

2. Crash when -mem is used. Fix: copy max_lh_slots to the inner tree.

• Tested on LG+C10+G4 (20 taxa, 5000 sites): EM now improves correctly, runs with -mem 4G and -mem 1G without crashing, and the non--mem and BFGS runs are unchanged.

[StefanFlaumberg]

Hi @thomaskf,

The bug causing the -mem crash has already been fixed by yet not merged #173 in a more safe and general way: the max_lh_slots member, if zero, is always computed by a initializeAllPartialLh() call. In ModelMixture::optimizeWithEM(), this call happens at the end of the across-submodel for-loop.
I think having max_lh_slots always checked and set automatically is a better solution than copying it manually because if a similar problem occurs in some similar place in the code where a temporary copy of the tree is used, it will be already solved.

Best regards,
Stefan

[thomaskf]

Thanks @StefanFlaumberg
Since the -mem issue is already addressed in your PR, there is no need to fix it here as well. I have therefore removed that fix, leaving only the fix for the EM optimisation issue.

[bqminh]

all good