fix: GET of a never-put memo index is an error (#47)

[jensens]

Fixes #47

A GET/BINGET/LONG_BINGET of a memo index that was never put yielded None when a higher index had been put before: memo_put grows the memo vector with placeholders, and a gap below the highest index read as one. CPython raises Memo value not found at index N.

What changes

The memo is a Vec<Option<PickleValue>>: growth fills with None, every store writes Some, and memo_get maps a None to the memo index N not found error the out-of-range case already raised. The dirty-flush sites (a memoized container mutated after the put) write Some. Option<PickleValue> uses the enum's niche, so the element stays 56 bytes and the vector does not grow; the FileStorage decode and PG pipeline A/B against main are within the noise floor.

No pickler writes such a stream; this is corruption handling, and the pre-scan optimisation (puts of indices never read are skipped) is unaffected because the gap error is path-independent.

Evidence

• 241 Rust and 403 Python tests; the tests cover the gap on the pre-scan path, on the fallback path (a trailing byte after STOP), in protocol 0 text opcodes, a dirty entry after a gap put, every Python entry point, and the memo shared across the two pickles of a ZODB record (a class pickle with its own puts makes index 0 valid in the state pickle).
• The 1,692-record sample is byte-identical to main.
• A fresh reviewer ran 80,000 random memo opcode streams against CPython: on every stream without MEMOIZE the branch now matches CPython's error decision exactly (832 streams that main decoded to a value are errors in both), and it reports the same index. The remaining divergence, MEMOIZE after a re-put of a filled slot, predates this branch (Decode regression since 1.6.0 (memo dirty-sync clones every container); pre-scan memo reads to skip unread puts #22) and is filed as a follow-up issue.

Assisted-by: Claude Fable 5.1