perf(cli): parallelize package scanning with rayon

[BryanFRD]

Package
Which package does this affect? [x] cli [ ] action [ ] schema

Problem / Motivation

In `src/monorepo.rs`, `run_release_logic` iterates over `config.packages` sequentially (line 438: `for (pkg_idx, pkg) in config.packages.iter().enumerate()`). For each package it:

1. Calls `crate::git::find_highest_semver_tag` (walks all tags via `tag_foreach`).
2. Calls `get_commits_since_last_stable_tag` / `get_commits_since_last_tag` (revwalks the entire commit history from HEAD).
3. Calls `read_version` (file I/O + parsing per format).

These are independent per-package: no shared mutable state, no inter-package ordering until the cascade phase that runs afterwards. On a 200-package monorepo (see `Fixtures/fixtures/examples/large-monorepo.json`) the wall time is dominated by the per-package revwalk, which is embarrassingly parallel.

Benchmarks (`Benchmarks` repo, fixture `mono-large-ferrflow-binary-check`) currently have absolute thresholds of 2000ms for check / 500ms for version on the large monorepo — there's headroom to claim a 4-8× speedup on multi-core CI runners.

Proposed solution

1. Add `rayon = "1"` to `cli` feature deps.
2. Extract the per-package analysis loop into a pure function: `(pkg, &config, &repo_root, &all_tags) -> PackagePlan { name, current, next, bump, commits, tag }`.
3. Wrap with `config.packages.par_iter().map(...).collect::<Result<Vec<_>>>()`.
4. Keep the write/commit/tag/hook side-effect phase sequential — it must remain deterministic.

`git2` repository handles are not Send, so each rayon thread opens its own `Repository` (cheap — discover only). Pre-compute `all_tags` once on the main thread before fanout, so the parallel work doesn't repeat `collect_all_tags`.

Alternatives considered

• `tokio` async — `git2` is sync, no benefit.
• Don't parallelize — fine for repos with <10 packages, but the benchmark fixture sized at 200 packages is unrealistic to keep serial.

Additional context

Severity: P2. Closes / aligns with #417 (wire benchmarks for regression tracking) — once parallelisation lands the benchmark numbers become the gate against future regressions.

[BryanFRD]

Deferring out of the audit batch (PR #435). The target loop in src/monorepo.rs:438 holds git2::Repository (which is not Send / Sync), mutates ~7 pieces of shared state inside the body (pkg_outputs, tags_to_create, hook_contexts, files_to_commit, bumped_names, any_bumped, json_packages), and uses ? to propagate errors mid-iteration. A correct rayon port needs per-thread Repository::open, a split into gather/decide/emit phases, and an error-aware combinator like try_fold — bigger than the "minimum change" this issue scoped. Will land in its own PR.