Expose TensorNetworks.Evaluator and typed evaluate_batch

[shinaoka]

Motivation

BubbleTeaCI Task9 migration exposed that repeated Tensor4all TreeTN/TensorTrain evaluation is too expensive for TCI-style sampling. Tensor4all.jl currently calls the C API in a way that rebuilds backend handles per evaluation call, and evaluate returns ComplexF64 even for real TreeTNs.

Proposed API

Expose a reusable evaluator wrapper after tensor4all-rs provides the C API:

ev = TensorNetworks.Evaluator(tt, site_indices)
values = ... # shape (n_site_indices, n_points), 1-based Julia indices
out = ev(values)

Return type should follow scalar kind:

• real TreeTN/TensorTrain -> Vector{Float64}
• complex TreeTN/TensorTrain -> Vector{ComplexF64}

Conventions

Keep Tensor4all.jl aligned with tensor4all-rs:

• batch input shape is (n_site_indices, n_points)
• the point/batch dimension is rightmost
• Julia wrapper converts 1-based input values to Rust's 0-based values

Python bindings may expose torch-style (batch, n_sites) separately; Julia should not change the core convention.

Acceptance criteria

• TensorNetworks.Evaluator(tt, indices) reuses backend evaluator state across calls.
• Real evaluation returns Vector{Float64} and scalar Float64 where applicable.
• Complex evaluation returns Vector{ComplexF64} / ComplexF64.
• Existing evaluate(tt, indices, values) remains compatible or has a documented migration path.
• Tests cover real and complex TreeTNs and repeated evaluator calls.

Upstream dependencies

• Add allocation-light TreeTN batched evaluator tensor4all-rs#464
• Expose reusable TreeTN evaluator through the C API tensor4all-rs#465
• Make TreeTN evaluation outputs typed by scalar kind tensor4all-rs#466