fix(fts): synchronize wildcard DF accumulation

[future3OOO]

Summary

• synchronize wildcard DF accumulation in MatchTermsVertexCompute
• share a single mutex across copy() worker instances
• remove the unsynchronized unordered_map write race in the wildcard FTS path

Root cause

MatchTermsVertexCompute stores std::unordered_map<offset_t, uint64_t>& resDfs and copy() returns new worker instances that share that same map. Ladybug runs those copies concurrently from VertexComputeTask::run(), so wildcard matches can execute resDfs[nodeIds[i].offset] = dfs[i]; from multiple worker threads at once.

Why this fix

This is the smallest correctness-first change that matches the demonstrated race without changing the existing VertexCompute API. A thread-local merge design may recover more parallelism later, but it needs extra merge plumbing that does not exist on the current worker interface.

Validation

• source-level review against current main (160c616cdb0688fd00fdcb4686bb269801a374d4)
• workload-level reproduction was previously observed from GitNexus wildcard augment calls with native frames through MatchTermsVertexCompute, VertexComputeTask::run(), and std::__detail::_Map_base<...>::operator[]
• this environment does not currently have a compiler toolchain (gcc/g++/clang absent), so I could not rebuild the extension locally before opening this draft PR

Follow-up

If maintainers want the higher-throughput design, I can rework this toward worker-local accumulation plus post-scan merge once the merge point is agreed.