Iterator-based sweep solvers

[lkdvos]

Description

Restructures MPSKit's sweep-based solvers as iterators with a common shape. This is a pure refactor: results, logging and returned values are unchanged.

Every converted solver now has a state struct, and Base.iterate runs sweep!(it, state, Val(:right), iter) followed by sweep!(it, state, Val(:left), iter), then any finalize or error step. Finite sweeps take their sites from _sweep_ranges. Converted:

• finite find_groundstate with DMRG/DMRG2 (DMRGState, sweep_converged)
• finite approximate! with DMRG/DMRG2
• propagator with DynamicalDMRG (NaiveInvert, Jeckelmann)
• multiline approximate! and leading_boundary with IDMRG/IDMRG2
• time_evolve/time_evolve!: one iteration is one time step plus finalize

Finite TDVP/TDVP2 timestep! is now a per-site local_update! plus a plain sweep, following BUG.

Smaller cleanups:

• IterLog(::Algorithm) names logs after the algorithm type.
• The redundant = it.state defaults on Base.iterate are removed; iterativesolvers.jl already provides the one-argument method.
• state arguments are now typed, which avoids method ambiguities with the generic time_evolve iterator.

This gives the finite environment cache from ld-envs per-site hook points in DMRG, TDVP/TDVP2, BUG and dynamical DMRG. ld-envs will be rebased onto this branch; the rebase applies cleanly. It is also a step towards migrating these solvers to AlgorithmsInterface.jl, which will be a separate PR.

Verification

• Equivalence against main: equivalence scripts ran the same cases on main and on this branch with fixed seeds, comparing with isequal. Compared: the raw MPS tensors, AlgorithmInfo fields, energies and returned values, every iteration log record and warning, and the arguments of finalize calls. All 168 cases are identical: 28 DMRG/approximate, 8 dynamical DMRG, 18 IDMRG and 114 TDVP/BUG/time_evolve.
  • Coverage includes converging and maxiter-limited runs, maxiter = 0, CBE and DMRG3S, nearest-neighbour and long-range models, real-valued initial states, time-dependent operators, and empty or single-point t_span.
  • The only differences were TensorKit cache-miss debug messages, which depend on whichever Julia session ran first, not on this change.
• Tests: the groundstate, approximate, dynamical DMRG, statmech, timeevolution/*, states/*, fidelity susceptibility and anyonic braiding test files pass.
• Ambiguities: Test.detect_ambiguities(MPSKit) is unchanged from main (135).

Checklist

• Tests pass locally (relevant subset, see above)
• Documentation updated, if this PR changes public API: no public API change
• Runic formatter is run
• Changelog entry added under [Unreleased] in docs/src/changelog.md

🤖 Generated with Claude Code