…appers

Followup to #354. Removes the deprecated/experimental `AbstractITensorNetwork`-
and `AbstractIndsNetwork`-side per-edge index wrappers in favor of the kept
`linkinds(tn, edge)` accessor and explicit `intersect`/`setdiff` calls on
raw `inds(...)` at call sites.

## Deletions

- `commoninds(tn::AbstractITensorNetwork, edge)` — was a thin alias for
  `commoninds(tn[src(e)], tn[dst(e)])`. The kept `linkinds(tn, edge)`
  accessor returns the same thing. The two were aliases anyway.
- `uniqueinds(tn::AbstractITensorNetwork, edge::AbstractEdge)` and the
  `Pair` overload — was `uniqueinds(tn[src(e)], tn[dst(e)])`.
- `hascommoninds(tn::AbstractITensorNetwork, edge::AbstractEdge|Pair)` —
  was `hascommoninds(tn[src(e)], tn[dst(e)])`.
- `uniqueinds(is::AbstractIndsNetwork, edge::AbstractEdge|Pair)` — walked
  the IndsNetwork to collect site inds at `src(edge)` plus link inds on all
  other edges incident to `src(edge)`. The single internal caller (in
  `ttn(::ITensor, ::IndsNetwork)`) inlines a semantically-equivalent rewrite
  using the already-built `tn`.

## Reimplementation

`linkinds(tn::AbstractITensorNetwork, edge)` is kept (parallel to the
also-kept `siteinds(tn, vertex)`) and reimplemented natively as
`intersect(inds(tn[src(e)]), inds(tn[dst(e)]))` instead of delegating to
the deleted `commoninds(tn, edge)` wrapper.

## Call site rewrites

- `weights(::AbstractITensorNetwork)` (`abstractitensornetwork.jl:34`) —
  `commoninds(graph, e)` → `linkinds(graph, e)`. The misleadingly-named
  `graph` argument is actually an `AbstractITensorNetwork`; this site
  was relying on the deleted wrapper.
- `IndsNetwork(tn)` and 1-arg `linkinds(tn)` (`abstractitensornetwork.jl`
  lines 249, 267) — `commoninds(tn, e)` → `linkinds(tn, e)`.
- `fix_edges!`, `union(tn1, tn2)` (lines 138, 183) — `hascommoninds(tn, e)`
  → `!isempty(linkinds(tn, e))`.
- `insert_linkinds` (line 848) — `!hascommoninds(tn, e)` →
  `isempty(linkinds(tn, e))`.
- `LinearAlgebra.svd`/`qr`/`factorize`, `gauge_edge`, `_truncate_edge`
  (5 sites) — `uniqueinds(tn, edge)` →
  `setdiff(inds(tn[src(edge)]), inds(tn[dst(edge)]))`.
- `ttn(::ITensor, ::IndsNetwork)` loop in `treetensornetwork.jl:233` —
  `uniqueinds(is, e)` → `setdiff(inds(a), inds(tn[dst(e)]))`. Equivalent
  given the post-order-DFS invariant: `tn[dst(e)]` retains its original
  inds (parent vertices aren't processed as `src` until after their
  subtree), so the only inds shared with the running accumulator `a` are
  the link inds on edge `e`.
- 2 sites in `test/test_itensornetwork.jl` "Index access" testset.

## Doc sync

Removes the corresponding entries from `docs/src/deprecated_methods.md`
(per-edge `commoninds` / `uniqueinds` / `linkinds` block) and
`docs/src/experimental_methods.md` (IndsNetwork-side `uniqueinds`).
Drops dead imports (`hascommoninds`, `unioninds`, `uniqueinds` from
`abstractitensornetwork.jl`; `ITensors`, `IndexSet`, `uniqueinds` from
`abstractindsnetwork.jl`; `hascommoninds`, `uniqueinds` from
`test/test_itensornetwork.jl`).

## Version bump

`0.20.0 → 0.21.0` (breaking — public methods removed). The pin in
`test/`, `examples/`, and `docs/` `Project.toml` is bumped from `"0.20"`
to `"0.21"` to match.

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

…ope was lost

Continuing the precompile fix from the previous commit. Several files
across `src/` call bare `hascommoninds(t1::ITensor, t2::ITensor)` or
`uniqueinds(t1::ITensor, t2::ITensor)` (both ITensors functions on
ITensor pairs — these are legitimate, not the deleted network+edge
wrappers) but don't import those names themselves. They were relying on
`abstractitensornetwork.jl`'s top-level `using ITensors: ..., hascommoninds,
unioninds, uniqueinds, ...` to bring those names into the
`ITensorNetworks` module namespace via transitive scope.

The wrapper deletion in this PR removed those imports from
`abstractitensornetwork.jl` (since the wrappers it defined were the only
local users), which in turn breaks the transitive scope downstream files
were depending on. Cold precompile fails; Revise's incremental updates
mask the problem locally.

Files updated (each gets only the names it actually uses):

- `src/apply.jl` — adds `hascommoninds` to the existing
  `using ITensors: ...` line.
- `src/solvers/subspace/densitymatrix.jl` — adds
  `using ITensors: hascommoninds, uniqueinds`.
- `src/solvers/applyexp.jl` — adds `using ITensors: uniqueinds`.
- `src/treetensornetworks/projttns/projttn.jl` — adds `hascommoninds`
  to the existing `using ITensors: ITensor` line.
- `src/treetensornetworks/projttns/projouterprodttn.jl` — same.

`src/solvers/subspace/ortho_subspace.jl` already does bare
`using ITensors`, which brings everything into scope, so it's unaffected.

These imports made the transitive dependency that already existed
explicit. No semantic changes.

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