fix: replace diskcache with atomic file-per-key cache

[shntnu]

Follows up on #128, which introduced diskcache to fix write corruption from parallel workers. That fix was correct for the write-race problem, but diskcache uses SQLite internally, and SQLite doesn't scale under heavy concurrent access.

Problem

With 200+ parallel Snakemake jobs (JUMP Cell Painting pipeline), the shared SQLite database becomes a bottleneck. We hit three failure modes, each uncovered after fixing the previous one:

1. sqlite3.OperationalError: database is locked during Cache.__init__ - the constructor runs PRAGMA and INSERT INTO Settings, which are writes. 200 processes deadlock on init.
2. After adding retry-with-backoff on the constructor: diskcache.core.Timeout on cache.set() inside multiprocessing workers - steady-state write contention.
3. After wrapping get/set in retries too: it worked, but fragile. Every cache operation needed error handling for a problem that shouldn't exist.

SQLite is single-writer. WAL mode (which diskcache enables) allows concurrent readers, but every Cache() constructor is a writer. No timeout or retry configuration fixes 200 concurrent writer-on-open sessions - it's architectural.

Fix

Replace diskcache with file-per-key caching using atomic writes:

• Read: np.load(path) - no lock, no coordination, scales to any number of processes
• Write: tempfile.mkstemp + np.save(fd) + os.replace(tmp, path) - os.replace is atomic on POSIX (same filesystem), so readers see either the complete old file or the complete new file, never a partial write

This addresses the original corruption from #128 (atomic writes prevent truncation/partial reads) without introducing a shared bottleneck (no SQLite, no locks).

What changed

• _cache_read / _cache_write: two small helpers, stdlib only
• null_dist_cached and get_null_dists: use file helpers instead of diskcache.Cache
• Remove diskcache from dependencies
• New test: test_null_dist_cached_corrupt (verifies graceful recovery from corrupted cache files)

Test plan

• 68/68 unit tests pass (including parallel stress test from fix: atomic cache writes to prevent corruption from parallel workers #128)
• Full JUMP pipeline: 124 parallel copairs jobs, cold cache, zero failures
• Previous pipeline runs with diskcache: 222 failures, then 202, then 101 (each fix uncovered the next SQLite contention point)