Cache TagSet("Site") in v3's apply_mpo() to speed up KPM and other MPO-heavy paths

[joselado]

Summary

• mpscpp3/chain_session.h's apply_mpo() (both overloads) re-parsed TagSet("Site") from scratch on every call via out.noPrime(TagSet("Site")). apply_mpo() is the single most-invoked MPO-application helper in the file: every Chebyshev moment of KPM's dynamical-correlator recursion (kpm_moments_full/kpm_moments_accelerated) calls it once or twice, and it also backs time evolution, the generalized-eigenproblem solver, and CVM's bicstab solver.
• The exact same class of overhead (TagSet construction re-parsing its string argument every call) was already profiled with valgrind --tool=callgrind and fixed elsewhere in this file (arnoldi_smallest_real, idmrg_extend_HL/idmrg_extend_HR, PR Wire ITensor v3 C++ iDMRG backend into infinitechain.py #87's follow-up) via a function-local static const TagSet — but that pass never touched apply_mpo() itself. This applies the identical, already-validated pattern there. Zero semantic change (same parsed tag value either way).
• Measured directly on the actual KPM hot loop: ~2-4% faster when bond dimension/moment cost is small enough that applyMPO's own SVD work doesn't dominate the per-call fixed overhead; negligible when it does (unlike the iDMRG Arnoldi case's 6-8%, since that loop spends proportionally more time in cheap per-step ITensor operations relative to the parsing cost). v2 is unaffected — it predates the Tags system.
• Also fixed examples/dynamical_correlator/dynamical_correlator_v2_VS_v3/main.py, found broken while benchmarking this change: it used a stale i=i,j=j,name="ZZ" shorthand that get_dynamical_correlator no longer accepts (name= must be a pair of MultiOperators), so it raised RuntimeError: No active exception to reraise on both backends. Replaced with name=(sc.Sz[i],sc.Sz[j]), the same convention tests/test_dynamical_correlator.py already uses.

Test plan

• pytest tests/test_dynamical_correlator.py -k "kpm or cvm" (parametrized over itensor_version incl. 3) passes before and after the change
• Full suite: pytest tests — 268 passed, 9 skipped, unchanged from baseline
• Benchmarked before/after with a standalone script exercising get_dynamical_correlator(mode="DMRG", submode="KPM") at multiple chain sizes/moment counts, built via the fast worktree-local mpscpp3 rebuild trick (reuse libitensor.a, just make pybind)
• Fixed example re-run end-to-end against both real compiled backends (mpscpp2 + mpscpp3, not the ED fallback), stable timings across repeated runs

🤖 Generated with Claude Code

https://claude.ai/code/session_01DoZKWzhMo7qKoiP12gbcvv