Consolidate edge factorization API on add_edges / qr / qr! - #363

Replaces the manual per-edge loops in `add(tn1, tn2)` and the tebd
initial-state setup with `add_edges!`, which dispatches per-edge to
the reformed `add_edge!` (ensures graph edge + link inds).

Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

Rewrite the only caller, `TreeTensorNetwork(::ITensor, ::IndsNetwork)`,
to build up the vertex-tensor `Dict` directly during the post-order DFS
instead of scaffolding a placeholder `ITensorNetwork` upfront. The
scaffold tensors' link inds (filled in by `insert_linkinds`) never
participated in the factorization — they were stripped by
`setdiff(inds(a), inds(tn[dst(e)]))` and then overwritten by the
`tn[src(e)] = a_l` writes anyway. The literal effective behavior was
`left_inds = setdiff(inds(a), site_inds[dst(e)])`, which is what the
new code computes directly.

With the IndsNetwork-form caller gone, the function (and its
`experimental_methods.md` entry) follow.

This does not change the algorithm's observable behavior — including
the existing quirk that the resulting TTN does not put each
`site_inds[v]` on its own vertex (the first vertex processed in
post-order absorbs every site ind that isn't `dst(e)`'s, etc.). That
structural issue is real but out of scope here; tracked separately.

Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

The constructor had no in-src callers and only two test-only callers
(`test/test_ttns.jl`'s 'Construct TTN from ITensor or Array' and
`test/test_ttno.jl`'s 'Construct TTN operator from ITensor or Array'),
each of which only verified that `contract(ttn) ≈ original_dense`.
Tracing the algorithm showed it was producing factorizations that
contract back to the original but do not lay out site indices on their
designated vertices — the first vertex processed in post-order absorbs
every site index that isn't `dst(e)`'s, leaving later vertices without
their own site indices. The round-trip tests passed despite this
because the misplaced site indices and trivial links still contract
correctly.

Drop the constructor and its two test callers rather than rewrite the
algorithm: nothing outside the tests was relying on it, and a correct
replacement can be designed if a real use case appears.

Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

… ctor

The 'From a dense ITensor' section documented
`TreeTensorNetwork(::ITensor, ::IndsNetwork)`, which was deleted in
f704f66. The `@example` block would have crashed the doc build.

`add(tn1, tn2)` and `tebd` were both calling `add_edges!` on local
copies (already returned by `combine_linkinds` or a preceding
`copy(ψ)`). Use `add_edges` (non-mutating) directly: same semantics,
clearer functional style, and the explicit `copy(ψ)` in `tebd` is no
longer needed.

Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

The only caller — the multi-edge collapse step inside `add(tn1, tn2)` —
is replaced by an inline loop calling `factorize_edge!(qr, tn, e)` on
multi-edges. `factorize_edge!(qr, ...)` produces the same single-shared-
Index end state as `combine_linkinds`, plus a free rank truncation to
the actual rank of `tn[src]` on the (left-inds | multi-link)
bipartition. With no callers left, drop `combine_linkinds`,
`linkinds_combiners`, and `is_multi_edge`, and remove the
`combine_linkinds` entries from `docs/src/experimental_methods.md`.
The `combiner` import in `abstractitensornetwork.jl` is no longer
needed either.

Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

- Drop the `_ITensorCollection` Union from the `ITensorNetwork` constructor
  signatures and the in-file comments referencing the long-removed
  function-callback ctor's dispatch loop (those constructors went away in
  #356). The `Vector{ITensor}` / `Dict{V, ITensor}` / `Dictionary{V, ITensor}`
  shapes still flow through the same body via duck-typed `keys` /
  `keytype` / `tensors[v]`.
- Use `LinearAlgebra.qr` directly inside `qr!` (instead of
  `factorize(...; ortho = "left")`), and reorder the residual contraction
  to `R * tn[dst(edge)]` so the expression reads left-to-right as
  "absorb the residual into the destination tensor". Drop the now-unused
  `factorize` import.
- Refresh the stale `gauge` entry in `docs/src/experimental_methods.md` —
  the `alg::Algorithm` positional and `kwargs...` were dropped from
  `gauge` earlier in this PR.

Plain `tn[v] = ...` writes go through `set_vertex_data!`, which needs
to keep the graph edges incident to `v` in sync with the shared indices
of the new tensor. Restore `fix_edges!` for that purpose, keyed off the
non-`@preserve_graph` path; the two callsites that now maintain the
invariant explicitly (the tensors-only `ITensorNetwork{V}` constructor
and `_add_edge!` in `test/utils.jl`) bypass `set_vertex_data!` via
`@preserve_graph` and don't pay the cost. Plan is to drop this helper
once the invariant becomes structural — e.g., once an `ITensorNetwork`
reverse Index map makes auto-reconciliation either unnecessary or a
constant-time lookup.