Feature request: `miniter` (minimum sweeps) for DMRG/DMRG2

[lkdvos]

Feature request: miniter (minimum sweeps) for DMRG/DMRG2

Motivation

The DMRG/DMRG2 stop test now uses the Galerkin error as the convergence measure (previously the
overlap-based ϵ_conv = 1 - |⟨old|new⟩|).
The Galerkin criterion can be near-zero already on the first sweep when the initial guess is
close to an eigenvector, so the optimizer declares convergence and returns immediately — even
though more sweeps would move the state meaningfully.

This surfaced in FiniteExcited: with an unperturbed ground-state initial guess, the first
excited-state search converged instantly on the ground state instead of climbing to the excited
state. We worked around it there by perturbing the initial guess, but a general safeguard is
still desirable.

Proposal

Add a miniter::Int field to the ground-state sweep algorithms (at least DMRG and DMRG2,
possibly VUMPS/IDMRG for consistency), defaulting to 1 (no behavior change). Gate the
convergence break on it:

if ϵ <= alg.tol && iter >= alg.miniter
    ...
    break
end

Callers that want to force a minimum number of sweeps (e.g. FiniteExcited, or any run starting
from a near-eigenvector guess) can then set miniter > 1.

Notes / open questions

• Default should be 1 to preserve current behavior; callers opt in to a higher value.
• Decide whether FiniteExcited should bump its gsalg's miniter automatically (e.g. via
  Accessors.@set) or leave it entirely to the user.
• Scope: DMRG/DMRG2 only, or all iterative ground-state solvers.

Deferred from the change that switched DMRG/DMRG2 convergence to the Galerkin error.