v0.1: dataset fetch + sha256 verification + cache resolution

[protosphinx]

Stacked on top of #2 (end-to-end loop). Merge #2 first; then this rebases trivially onto main.

Summary

• Ships the v0.1 machinery: pm-bench fetch <name> [--pin], sha256 verification, cache directory resolution.
• Per-dataset hash-pinning is now a one-time manual step (download from 4TU/Mendeley → place in cache → run --pin → PR the patch). Everything downstream is automated.

What's new

• pm_bench/cache.py — cache root from $PM_BENCH_CACHE (else ~/.cache/pm-bench/), per-dataset path with the right extension by format. Synthetic datasets are rejected (generated on demand).
• pm_bench/fetch.py — ensure_cached(dataset) covers all four cases:
  1. cached + hash matches → return path
  2. cached + hash mismatch → HashMismatchError (loud, refuses silent proceed)
  3. cached + registry hash unset → returns actual hash (caller can --pin it)
  4. not cached → auto-download if URL set, else ManualFetchRequired with the precise landing URL and on-disk path the user needs to fill in
     Atomic .part-then-rename writes; streams in 1 MiB chunks.
• CLI: pm-bench fetch <name> [--pin] — prints status; with --pin against an unpinned cached file, emits a pasteable registry.yml sha256 patch.
• 13 new tests (test_cache.py, test_fetch.py). 37 total.

Why this matters

• v0.1's blocker was always the 4TU TOS — we can't automate the download. This PR makes that the only manual step. Once a contributor accepts the TOS, downloads, and runs --pin, the registry update is a trivial PR; the rest of the toolchain (parse → split → prefixes → predict → score) is wired.
• Unblocks gnn's v0.5 milestone the moment any single dataset is pinned.

Smoke

$ pm-bench fetch synthetic-toy
synthetic-toy: generated on demand, no fetch needed

$ pm-bench fetch bpi2020
bpi2020: no download_url (TOS-gated). Visit https://data.4tu.nl/...,
  accept the terms, and save the archive to ~/.cache/pm-bench/bpi2020.xes.gz.
  Then re-run `pm-bench fetch bpi2020 --pin` ...

# (manual download)

$ pm-bench fetch bpi2020 --pin
bpi2020: cached at ~/.cache/pm-bench/bpi2020.xes.gz (unpinned)
  sha256: <hex>

  # paste under the matching dataset entry in pm_bench/registry.yml:
    - name: bpi2020
      sha256: <hex>

Test plan

• pytest -q — 37 passed (was 24 on PR End-to-end loop on synthetic-toy: prefixes + predict + score #2)
• ruff check pm_bench tests — clean
• Manual CLI smoke for synthetic-toy, manual-fetch error path, and --pin against a planted file

Roadmap impact

• v0.1 marked 🟡 in README/GOALS (machinery shipped, per-dataset pins pending). The next concrete step is per-dataset PRs that pin sha256, plus wiring an XES parser to _load_events (likely behind a [bpi] extra to keep the base install light).

[protosphinx]

Merged into main as part of the audit-cleanup stack (commit 9c00b47). The full content of this PR is now on main.