perf: Batch-construct vertex sequence lists via create_vs_list()

[schochastics]

Split 2 of 4 out of #2696, per #2696 (comment) — this is the requested lapply() -> create_vs_list() step.

Base: #2886 (review that first; this PR's diff is only the top commit). Next in the stack: #2888.

What changed

37 call sites built lists of vertex sequences with

lapply(res, unsafe_create_vs, graph = graph, verts = V(graph))

which re-reads the graph reference, the graph id and the name source out of verts on every single object. create_vs_list(graph, idx_list) hoists all of that per-graph work out of the loop, so each sequence costs one as.integer(), one name subset and one attributes<-.

• The VERTEXSET_LIST OUTCONV template in tools/stimulus/types-RR.yaml now emits create_vs_list(graph, res), and the generated R/aaa-*.R files are updated to match (27 sites).
• Hand-written call sites updated in cliques.R, cohesive.blocks.R, components.R, conversion.R, interface.R, paths.R, structural-properties.R (10 sites).
• unsafe_create_vs() stays as the single-object form.

max_cliques(sample_gnp(200, 0.16)) on a named graph: 7.8 ms -> 5.6 ms (24.7 ms on main).

Why it matters beyond the numbers

This is the enabling refactor. It gives "turn this list of ID vectors into a list of vertex sequences" a single named entry point, so the construction loop can later be moved wholesale — into C in #2888, and behind lazy names in #2696 — without touching any of the 37 call sites again. Both of those PRs change only the body of this one function.

Scope

Edge sequences are deliberately left alone: simple_es_index() already propagates the shared weak reference from a single E(graph), so unsafe_create_es() does not pay the per-object cost. An create_es_list() batch form remains a follow-up.

Correctness

Output, names, NA handling and graph recovery are unchanged — this is a pure refactor of where the per-graph lookups happen.

Full suite: FAIL 0 | WARN 0 | SKIP 7 | PASS 9299.

[github-actions]

This is how benchmark results would change (along with a 95% confidence interval in relative change) if dec29d7 is merged into perf/vs-shared-weakref:

• ✔️E_named: 211ms -> 214ms [-0.56%, +3.17%]
• ✔️V_named: 428ms -> 429ms [-0.05%, +0.66%]
• ✔️as_adjacency_matrix: 241ms -> 240ms [-0.95%, +0.1%]
• ✔️as_biadjacency_matrix: 256ms -> 256ms [-0.5%, +0.64%]
• ✔️as_data_frame_both: 261ms -> 261ms [-0.84%, +0.51%]
• ✔️as_long_data_frame: 208ms -> 208ms [-0.81%, +0.39%]
• ✔️es_attr_filter: 244ms -> 243ms [-0.64%, +0.01%]
• ✔️graph_from_adjacency_matrix: 271ms -> 271ms [-0.42%, +0.45%]
• ✔️graph_from_data_frame: 273ms -> 274ms [-0.32%, +0.8%]
• ✔️head_of_named: 182ms -> 182ms [-0.72%, +0.94%]
• 🚀max_cliques_named: 68.6ms -> 58.4ms [-15.4%, -14.33%]
• ✔️vs_attr_filter: 331ms -> 330ms [-0.86%, +0.45%]
• ✔️vs_by_name: 311ms -> 311ms [-0.49%, +0.32%]
  Further explanation regarding interpretation and methodology can be found in the documentation.