Replace pass-through `...` with dedicated `options` arguments

[krlmlr]

In the 3.0.0 argument-migration program (#2757), ... becomes a pure
keyword-only separator: the generated ARG_HANDLE blocks own it to
recover legacy calls, and functions without recovery enforce
check_dots_empty(). Functions that forward their dots somewhere
else conflict with that design — and they already have the classic
dots problems today:

• typos and stale arguments vanish silently — worse, three live
  functions forward dots into a callee that ignores them entirely, so
  everything passed there vanishes (see traps below);
• no check_dots_empty() is possible, and the migration machinery
  cannot distinguish "argument for the callee" from "legacy positional
  argument to recover";
• the set of accepted names is invisible in the signature and docs.

Proposal (per maintainer direction): each forwarding function gets
a dedicated options argument with its own logic — a validated named
list (constructor helpers where a family warrants them, in the spirit
of the existing options = arpack_defaults() pattern), unknown-name
errors instead of silent swallowing, and ... reserved for the
separator + recovery like everywhere else.

Affected functions (live API; complete inventory)

Layout family — forwards to a chosen layout function (21):
layout_nicely(), add_layout_(), layout_components(),
tk_reshape(), and the 17 deferred-spec wrappers
as_bipartite(), as_star(), as_tree(), in_circle(), nicely(),
on_grid(), on_sphere(), randomly(), with_dh(), with_drl(),
with_fr(), with_gem(), with_graphopt(), with_kk(),
with_lgl(), with_mds(), with_sugiyama().
Natural shape: layout_options(...) or mirroring the target's
formals in the wrapper.

Constructor-spec sugar — forwards into make_/sample_ (8):
chung_lu(), degseq(), tree(), directed_graph(),
undirected_graph(), from_literal(), from_data_frame(),
from_edgelist().

Isomorphism family — dots depend on method (6):
isomorphic(), subgraph_isomorphic(), count_isomorphisms(),
count_subgraph_isomorphisms(), isomorphisms(),
subgraph_isomorphisms().
Strongest case for "options with its own logic": the valid option set
is a function of the chosen method (vf2 / bliss / lad).

Foreign I/O — dots depend on format (2):
read_graph(), write_graph() (per-format readers/writers).
Same shape as the isomorphism family.

Statistics / conversion / misc (6):
component_distribution(), degree_distribution(),
fit_power_law() (only its "R.mle" branch forwards),
scan_stat() (→ local_scan()), print_all() (→ print options),
as.matrix.igraph() (→ as_adjacency_matrix()/as_edgelist()).

Plotting S3 (8): plot.igraph(), rglplot.igraph(), plot.sir(),
plot.communities(), plot.cohesiveBlocks(), plot_hierarchy(),
plot_dendrogram.communities(), plot_dendrogram.igraphHRG().
S3 methods must keep ... for the generic contract; the question is
whether they additionally gain an options/params argument so the
graphical parameters stop hiding in dots.

30 deprecated shims also forward dots; per policy they are never
touched and simply follow (or outlive) their targets.

Traps found while inventorying (bugs today)

• chung_lu() forwards ... into sample_chung_lu(), which has
  ... before its own keyword args and ignores it — anything a
  user passes there vanishes silently.
• minimum.spanning.tree() → mst() and
  graph.cohesion() → cohesion.igraph() swallow dots the same way.
• make_graph()'s numeric branch does
  do.call(old_graph, list(edges, ...)), so unnamed extra dots can
  positionally bind the internal n/directed parameters.

Explicitly out of scope (dots are the payload, not options)

• Variadic-data functions where dots are the data (30):
  make_graph(), graph_from_literal(), edge()/vertex()/path(),
  add_edges()/add_vertices() (attributes), set_vertex_attrs(),
  disjoint_union(), the union()/intersection()/difference()
  families (dots-only dispatch), layout_() modifiers,
  igraph_options(), with_vertex_()/with_edge_()/with_graph_()
  (NSE), c()/[ methods on vs/es.
• User-callback forwarders where dots belong to FUN:
  local_scan(), head_print(), indent_print().
• The 17 functions already enforcing check_dots_empty(), and pure
  separators — already correct.

Questions for maintainers

• One generic options = list() everywhere, or per-family
  constructor helpers (layout_options(), per-method isomorphism
  options, per-format I/O options) with validation?
• For the deferred layout sugar (with_fr() etc.): options list,
  or mirror the target function's formals outright?
• Plotting: keep dots-only (S3 reality), or add a params/
  options argument alongside the contractual ...?
• Transition: recover dots-passed options via the ARG_HANDLE
  machinery with a soft deprecation ("passing layout parameters
  through ... is deprecated, use options ="), so old code
  keeps working through 3.x?
• Fix the three silent-swallowing traps above ahead of the
  redesign (they are user-visible bugs regardless)?

Related: #2757 (program), #2780 (@param ... docs), #2788 (renames),
#2789 (defaults). No existing issue covered this space, hence no
Closes lines.