perf(pr)!: batch prune's merged-PR lookup into one gh call

[lettertwo]

workon prune spawned one gh pr list --head <branch> --state merged per signal-less row to find the PrMerged signal. Since #99 added no-worktree branches to prune's scope, that cost scales with local branch count: on a repo with 80 local branches, a dry run made 79 sequential gh calls at ~0.4s each and took ~32s.

This PR replaces the per-branch lookup with one gh api graphql request that carries an alias per branch (up to 100 per request). The same dry run now takes 1.73s and makes one api graphql call, with identical output.

• find_merged_pr in the workon lib becomes find_merged_prs(&[&str]) -> Result<HashMap<String, MergedPr>>, which is why the commit is marked breaking.
• Each alias asks for pullRequests(headRefName:, states:MERGED, first:1) ordered by CREATED_AT DESC, matching the newest-first PR gh pr list returned. Branch names go in as -f variables, never into the query text.
• gh resolves {owner}/{repo} the same way gh pr list resolves its repo.
• The 3-consecutive-failures bound is gone. A failed request now degrades every row to no signal at once, with no warning, which keeps prune's under-report-only contract.
• The prune test stubs answer api graphql instead of pr list. The consecutive-failure tests are replaced by a test asserting one call covers every worktree and branch row, and a test asserting a failed call degrades quietly.