Referenda benchmarks

[l0r1s]

Referenda & multi-collective: split hook traits and add benchmarks

Summary

Adds proper benchmarks to pallet-referenda and pallet-multi-collective,
and reshapes the cross-pallet hook surface (poll hooks, members hook) so
each notification carries its own weight bound. Both pallets' extrinsics
are now weighted from real benchmarks instead of dev_mode's zero
defaults.

Motivation

Both pallet-referenda and pallet-multi-collective were running with
dev_mode enabled, silently zero-weighting every extrinsic. Several of
their extrinsics fan out into hooks living in other pallets (the voting
backend, the vote-cleanup hook), so even benchmarking each pallet's own
logic isn't enough — the static charge needs to include each hook's
worst case too. To weight them honestly we needed the hook traits to
publish their weight, and we needed real benchmarks for the extrinsics
themselves.

Changes

pallet-referenda

• Split the bundled PollHooks into per-event traits
  (OnPollCreated, OnPollCompleted); folded on_tally_updated into
  Polls since it shares the same provider type.
• Each hook trait now exposes a worst-case weight() so callers can
  charge the right static budget upfront.
• Dropped dev_mode, added a WeightInfo trait, and benchmarked
  submit, kill, advance_referendum, and on_tally_updated.
  advance_referendum is benchmarked on the worst-case branch
  (approve-with-Review, which fires both hooks); other branches are
  bounded conservatively from above.

pallet-multi-collective

• Members storage is now kept sorted by AccountId. Every write path
  (add_member, remove_member, swap_member, set_members) uses
  binary_search plus try_insert/remove, and set_members diffs
  against the previous set via
  ChangeMembers::compute_members_diff_sorted. The sorted invariant
  bounds the per-call cost so the extrinsics are actually benchmarkable.
• Renamed reset_members → set_members and SetMembersOrigin →
  SetOrigin; the call takes the full member list and overwrites, so
  "set" is what it actually does.
• force_rotate now returns DispatchResult and pays the static
  worst-case instead of refunding via OnNewTerm's returned weight.
• Added a WeightInfo trait, benchmarked every extrinsic, and wired
  weights as
  WeightInfo::*().saturating_add(T::OnMembersChanged::weight()).

Cross-pallet hook traits

• Moved OnMembersChanged from pallet-multi-collective into
  subtensor-runtime-common::traits, alongside Polls,
  OnPollCreated, and OnPollCompleted.
• All four traits now expose a worst-case weight(). The runtime's
  GovernanceVoteCleanup (the OnMembersChanged impl that wipes
  signed-voting state for outgoing members) bounds itself by
  MaxMembers × MaxQueued × (per-poll storage cost + on_tally_updated_weight).

Runtime wiring

• BenchmarkHelper config fields on both pallets, wired against the
  existing governance tracks and the Proposers / Economic
  collectives so the per-pallet benchmarks have a non-rotatable
  collective to fill freely and a rotatable one for force_rotate.

[shamil-gadelshin]

Cargo clippy failed.