docs: zoning rules for arguments around the ellipsis

[krlmlr]

Central PR: moving optional arguments past the ellipsis, repo-wide

Coordination point for applying the argument-migration pattern
(#2684, #2686, #2743) across all exported functions with optional arguments.
After #2779 landed the registry-split logic on main, this PR is docs-only:
the design rationale in CONTRIBUTING.md
(section Argument Order and the Ellipsis),
with pointers from AGENTS.md, tools/README.md,
and tools/migrations/README.md.

The zoning rules (summary)

Head (before ..., positional):

1. All required arguments, data first.
   This includes functions with three or more required arguments:
   a required argument after ... would be keyword-only but still required,
   which hides it from the signature and worsens error messages.
2. At most a few defining optional arguments — those that say what
   is computed, not how:
   the objects operated on (v in degree()),
   model parameters (power, m in sample_pa()),
   or a variant-selecting enum (type in as_adjacency_matrix()).

Tail (after ..., keyword-only): everything else, roughly ordered as
interpretation modifiers (mode, directed, weights, loops) →
output shape (normalized, names, sparse) →
algorithm/tuning (algorithm, options, callbacks) →
deprecated arguments last.

Deliberate deviation from the tidyverse design guide:
dots-after-required
puts ... immediately after the required arguments.
We allow a small head of defining optional arguments because idiomatic
positional calls (degree(g, 1:3), sample_pa(100, 1.5)) are pervasive in
20 years of scripts, books, and teaching material; deprecating the idiomatic
form would create warning noise without a safety benefit.
The safety value of keyword-only arguments lies in the long tail of options.

Invariants of the sweep:

• No behavior changes: defaults stay untouched, no renames in this wave
  (renames remain expressible in the registry via the bare-symbol convention
  and can be layered per function family later).
• Old positional/abbreviated calls keep working through the generated
  ARG_HANDLE recovery, with a single lifecycle::deprecate_soft("3.0.0", …).
• Deprecated functions are not touched.

Topic PRs (226 functions, 21 PRs)

Each topic PR is a single commit on main (the split logic landed via #2779).

PR	topic	functions
#2759	structural-properties (incl. components, paths)	37
#2760	games	36
#2761	make	23
#2762	layout	17
#2766	conversion (incl. adjacency, incidence)	11
#2763	centrality	10
#2768	topology / isomorphism	10
#2765	community	9
#2767	flow (incl. cohesive blocks)	9
#2770	interface / iterators	8
#2771	hrg + embedding	7
#2777	plotting / tkplot / par	7
#2764	centralization	6
#2769	cliques	6
#2772	motifs + graphlets	6
#2776	similarity / efficiency / assortativity	6
#2778	misc (foreign, sir, coloring, …)	5
#2758	trees / decomposition	4
#2773	operators / rewire	4
#2775	walks + cycles	3
#2774	bipartite + degseq	2

Not in this sweep (and why)

• simplify(), as_undirected(), contract() — already migrated in refactor: snake_case attr-comb options & arguments, soft-deprecate dotted names #2742.
• Deprecated functions (erdos.renyi.game(), sample_bipartite(), …) — never changed.
• Replacement functions (E<-, vertex_attr<-, …) — value handling is special.
• Shape-API contract functions (shapes(), shape_noclip()) — positional calls are the contract.
• sample_correlated_gnp_pair() (p ~ permutation), lattice() (dimvector ~ dim) —
  head/recoverable prefix clashes the fix(migrations): guard against head/recoverable prefix clashes #2743 guard rightly rejects; need manual treatment.
• ~20 functions whose optional arguments are all defining (nothing to move).
• The trailing-pass-through-dots family (mst(), read_graph(), isomorphic(),
  local_scan(), power.law.fit(), layout_nicely(), …) — their ... forwards
  to other functions, which the recovery machinery cannot disambiguate;
  needs a dedicated design.