refactor: move optional arguments of operators functions behind the ellipsis

[krlmlr]

Part of the repo-wide ellipsis migration coordinated in #2757 —
see that PR for the full rationale (CONTRIBUTING.md, Argument Order and the Ellipsis).
Based directly on main (the registry split landed via #2779); a single topic commit.

What this does

Inserts ... between the defining arguments (head) and the optional
modifiers (tail) of 3 exported operators functions.
Arguments after ... become keyword-only.

• Legacy positional or abbreviated calls are recovered by the generated
  ARG_HANDLE block and emit a single lifecycle::deprecate_soft("3.0.0", …) —
  behavior is unchanged.
• No defaults change, no arguments are renamed, deprecated functions untouched.
• Registry: tools/migrations/operators.R; blocks regenerated via
  Rscript tools/generate-migrations.R (idempotent, CI-checked).
• Rd usage/arguments updated mechanically (see note below).

New signatures

function	head (positional)	keyword-only tail
complementer()	graph	loops
compose()	g1, g2	byname, graph.attr.comb, vertex.attr.comb, edge.attr.comb
each_edge()	prob	loops, multiple, mode

Placement rethink: keeping_degseq() was originally migrated here with an
empty head (all arguments keyword-only) — such signatures are now
disallowed (the head must never be empty; see the updated rule in #2757), so
it was dropped from this PR and keeps its original signature.

Notes for review

• The environment used to prepare this PR cannot install igraph.r2cdocs
  (GitHub API unreachable), so man/*.Rd files were updated mechanically
  (usage + ... argument item) instead of via devtools::document().
  R CMD check's usage↔formals validation passes; a follow-up
  devtools::document() run may reflow whitespace but should produce no
  semantic diff.
• Package tests that called these functions positionally were updated to
  named arguments (they are the same soft-deprecation user code will see).

[github-actions]

This is how benchmark results would change (along with a 95% confidence interval in relative change) if 40f194b is merged into main:

• ✔️as_adjacency_matrix: 817ms -> 814ms [-1.27%, +0.51%]
• ✔️as_biadjacency_matrix: 790ms -> 788ms [-1.2%, +0.69%]
• ✔️as_data_frame_both: 1.65ms -> 1.62ms [-6.32%, +1.99%]
• 🚀as_long_data_frame: 4.28ms -> 4.17ms [-4.83%, -0.29%]
• ✔️es_attr_filter: 2.96ms -> 2.98ms [-2.63%, +3.95%]
• ✔️graph_from_adjacency_matrix: 123ms -> 122ms [-1.7%, +0.59%]
• ✔️graph_from_data_frame: 3.5ms -> 3.53ms [-1.2%, +2.76%]
• ✔️vs_attr_filter: 1.68ms -> 1.69ms [-1.62%, +3.25%]
• ✔️vs_by_name: 1.11ms -> 1.09ms [-4.02%, +0.43%]
  Further explanation regarding interpretation and methodology can be found in the documentation.

[github-actions]

This is how benchmark results would change (along with a 95% confidence interval in relative change) if 76ee1bb is merged into main:

• ✔️as_adjacency_matrix: 801ms -> 797ms [-2.13%, +1.15%]
• ✔️as_biadjacency_matrix: 797ms -> 792ms [-1.58%, +0.12%]
• ✔️as_data_frame_both: 1.5ms -> 1.52ms [-1.41%, +3.68%]
• ✔️as_long_data_frame: 3.89ms -> 3.97ms [-3.27%, +7.71%]
• ✔️es_attr_filter: 2.97ms -> 2.92ms [-6.73%, +3.43%]
• ✔️graph_from_adjacency_matrix: 125ms -> 125ms [-1.55%, +2.45%]
• ✔️graph_from_data_frame: 3.37ms -> 3.39ms [-1.65%, +2.53%]
• ✔️vs_attr_filter: 1.67ms -> 1.63ms [-5.03%, +1.01%]
• ✔️vs_by_name: 1.05ms -> 1.04ms [-4.47%, +2.45%]
  Further explanation regarding interpretation and methodology can be found in the documentation.