complete per-command flags in the shell completions

[Kirananandan]

Closes #96.

Past the subcommand, the completion scripts only ever offered the five shared flags (-q --quiet --json --dry-run --help), so command-specific flags like restore --clean, diff --stat/--name-only, find --ids, save -n, or gc --keep never showed up on tab.

What changed

• _flags_for_commands() — a helper next to _subcommands() that walks each subparser's _actions off the argparse parser and maps every public subcommand to its option strings, so the map stays in sync as flags are added.
• Four small per-shell serializers (_bash_flagmap, _zsh_flagmap, _fish_flagmap, _pwsh_flagmap) render that map into each script's native form (bash/zsh associative array, fish complete lines, PowerShell hashtable), injected via a new __FLAGMAP__ placeholder.
• The dash branch of each of the four shell blocks now offers ${command}'s own flags instead of the shared five. Refs and file fallback are unchanged.
• Fish additionally keeps top-level flag completion (--version, --help) before a subcommand is chosen — the old unguarded global complete line offered those, and the per-command rewrite would otherwise have dropped them. Derived from the root parser via _root_flags().

Scope stayed in cli.py as the issue suggested, plus a docs note and a changelog entry.

Testing

• Added per-shell tests asserting each command's real flags land in the map and don't leak across commands (e.g. --keep doesn't appear under restore), plus a fish root-flag test.
• Full suite green (298 passed).
• Verified the generated bash and PowerShell scripts functionally on real shells: restore --<TAB> → --clean --only-missing ..., gc --<TAB> → --keep --keep-named --keep-within, and the hyphenated check-ignore --<TAB> → --exit-code all complete correctly. The PowerShell script also parses clean via the PS parser.