Fix --scfl gap detection for models with rate heterogeneity

[roblanf]

Summary

• Fixes --scfl giving incorrect results with any model using multiple rate categories (+G, +R, +I+G, +I+R, mixture models)
• The gap detection filter in computeSubtreeAncestralState() used a fixed threshold (sum > 1.0) that broke for ncat_mix > 1 and for deep subtrees with scaled partial likelihoods
• The fix uses scale_num to distinguish real data from gaps at internal nodes, handling both normal and SAFE_NUMERIC modes

Test plan

• Empirical validation on reporter's dataset (10 models × 4 methods)
• Simulation study (8-taxon tree, 1,100 runs across 4 models, 5 methods, 11 concordance levels, 5 gap levels)
• sCFL = sCFL --safe (max |diff| = 0.01)
• Single-category models unchanged (old == new)
• Multi-category models fixed (sN improved from ~1 to ~590 informative sites)

Validation scripts and full results: https://github.com/roblanf/iqtree3/tree/issue-137/tests/scfl_simulation

Full report of empirical and simulation test results: #137 (comment)

Fixes #137

🤖 Generated with Claude Code

[roblanf]

@bqminh can you take a look at this - would be great to get it into the latest binaries ASAP.

[bqminh]

looks good, thanks a lot for finding this out!