Mixlen refactoring: Make GHOST compatible with PMSF

[StefanFlaumberg]

This pull request implements the GHOST rate heterogeneity model for the PMSF model, a long-awaited feature requested several times before:  iqtree/iqtree2#391 (comment) and iqtree/iqtree2#459.

Implementation details:

• Fix a latent bug in the PhyloTreeMixlen::computeFuncDerv() function by switching the condition for using the base class-version function from if (cur_mixture < 0) to if (initializing_mixlen) and by asserting that cur_mixture is set. The rationale is that the base class-version function should be called only when the heterotachy model is not used yet, i.e. when initializing_mixlen is true. Otherwise, it would, in turn, call PhyloTree::computeLikelihoodDerv(Generic)SIMD() that would run in the mixlen mode, thus returning partial derivatives instead of the single ordinary derivative. Also, the unused code duplicating the one from phylokernelnew.h is now removed from this function.
• In the PhyloTree::computeLikelihoodDerv(Generic)SIMD() function, in the if (isMixlen()) branch, duplicate the if (SITE_MODEL) block of the else branch with the only difference that the length is considered category-specific here. Make the two code branches more similar.
• Change ->length to ->getLength(c) throughout the phylokernelnew.h code, most importantly in the if (SITE_MODEL) blocks, to ensure that all computations are compatible with the mixlen mode.
• For now, the PMSF model can be run only under the BFGS optimization of heterotachous branch lengths. Users running the default command with the EM algorithm will get the annoying error message advising switching to the BFGS algorithm. This inconvenience is intended to make the users informed that they have to run the optimization algorithm different from the one used in the original GHOST model paper. However, in another upcoming pull request, I am going to make the EM optimization work as well, so the restriction should be lifted soon.

A simple test to check the new likelihood calculations:
Run cmd1: iqtree3 --seed 123 -nt 1 -s aln.faa -m LG+F+H2 -optlen BFGS --keep-ident -pre test_LG_F_H
and cmd2: iqtree3 --seed 123 -nt 1 -s aln.faa -m LG+H2 -fs empty.sitefreq -optlen BFGS --keep-ident -pre test_LG_SSF_H,
where empty.sitefreq is literally an empty text file:

$ wc empty.sitefreq
0 0 0 empty.sitefreq

This two runs should have exactly the same likelihood, tree, and parameter estimates, since -fs empty.sitefreq activates a +SSF model that uses the +F state probability vector for each alignment pattern, which simulates the simpler +F model yet using site-specific computations.

[StefanFlaumberg]

@bqminh,
I was hoping it'd be included in the v3.1.3 release...

[bqminh]

I'm reviewing this now ... it's touching the kernel, so need be to thorough

[bqminh]

looks great. thanks. Suggestion: for future pull requests, better separate into smaller PRs, instead of combining many into one.

[bqminh]

It's now released in v3.1.3

[StefanFlaumberg]

It's now released in v3.1.3

@bqminh, thank you!

[StefanFlaumberg]

Hi @bqminh,

The v3.1.3 precompiled binaries exactly reproduce the behaviour of the v3.1.2 binaries when run with the PMSF+GHOST model: print out the wrong initial likelihood and immediately crash (thanks to @kr3118 for pointing this out to me).
I've tested only the iqtree-3.1.3-Linux-intel package, but others may have this problem as well. On the other hand, the binary compiled from the v3.1.3 source code runs fine. I guess you must somehow have added wrong binaries for the release.

Could you recompile and update the binaries if it is possible? Thanks!

Best,
Stefan