refactor: rename backend ref->scope, hard-remove legacy --name modifier syntax

[barrettruth]

Problem

Two related cleanups, no backwards compatibility kept (per AGENTS.md's
"Backwards compatibility" rule).

(1) The forge.Forge interface in lua/forge/types.lua declares
methods with scope?: forge.Scope, and the github backend already uses
scope. The gitlab and codeberg backends used ref as the local
parameter name, an unforced inconsistency that costs readability and
pattern-matching across backends. Issue #409.

(2) :Forge accepted a legacy --name / --name=value modifier
syntax in parallel with the canonical name=value form. The legacy
syntax was carried as documented "compatibility sugar", supported by a
parallel legacy_modifiers field on forge.CommandVerbDef /
forge.Command, a separate M.legacy_modifier_names public function,
dedicated branches in the parser and completion, and a separate
"legacy_flag" path in M.complete. None of the verbs in the families
table actually declared legacy_modifiers, so the field was dormant
production-wise — but the parsing/completion paths still accepted
--name=value for any modifier, providing a second way to do the same
thing.

Solution

(1) Mechanically rename ref -> scope across all method
signatures and bodies in lua/forge/backends/gitlab.lua and
lua/forge/backends/codeberg.lua, plus the local repo_arg(scope) /
project(scope) / hostname(scope) helpers. Surgical reverts for
'--ref' (CLI flag for glab ci list), entry.ref /
json.head.ref / json.base.ref (external API field names), and
normalize_run's local ref (a git ref like
refs/merge-requests/N/head, not a scope). Renamed gitlab.lua's
module import from local scope = require('forge.scope') to
local scope_mod = ... to free scope for parameter use and to match
the convention already in init.lua, resolve.lua, and log.lua.

(2) Hard-remove the legacy --name syntax. No deprecation, no
shims:

• forge.CommandVerbDef.legacy_modifiers and
  forge.Command.declared_legacy_modifiers /
  forge.Command.legacy_modifiers fields removed from
  lua/forge/types.lua.
• M.legacy_modifier_names removed from lua/forge/cmd.lua.
• token_is_modifier_like no longer treats ^-- as modifier-like;
  only name=value and bare flag tokens are considered.
• Parser body simplified: the if token:match('^--') branch and the
  parallel legacy_modifiers allowed-set are gone. Only name=value
  and bare flag tokens (when the modifier is declared kind = 'flag')
  parse.
• completion_state and filtered_modifier_completion_items lose
  the legacy branch and the legacy parameter.
• M.complete loses the legacy_flag block at the top and the two
  if arglead:match('^--') branches in family-slot and
  argument-slot completion.
• completion_policy.declares_modifier no longer falls back to
  legacy-modifier checks.
• Help text drops the "Compatibility sugar is still accepted ..."
  paragraph from the canonical-syntax section.
• :Forge pr create --draft --fill --web and
  :Forge issue create --blank --template=bug test cases updated to
  canonical draft fill web / blank template=bug.
• cmd.parse({ 'review', '42', '--wat=1' }) test updated to
  wat=1.

User-visible: any :Forge invocation that used --name / --name=value
must switch to name=value (or bare flag for draft / fill / web /
blank / all). Any tab-completion that previously offered --name
candidates after -- arglead no longer does. No migration window. The
canonical syntax has been the documented primary interface throughout.

just ci clean: nix fmt, stylua, biome, selene, lua-language-server
check, vimdoc-language-server, 633/0/0 busted.

Closes #409.