Refresh SizeMeasure support and make archived row decoding fallible

[Handy-caT]

Refreshes the original SizeMeasure work onto current main, modernizes dependencies, and closes the archived-row safe-API soundness issue.

Changes

• adds an overridable SizeMeasurable::default_aligned_size type-level helper
• adds generic primitive-array sizing while preserving the existing array alignment contract
• adds SizeMeasure support for non-generic fieldless enums, measured from the explicitly named rkyv archived representation
• rejects payload-bearing enums instead of guessing a data-dependent size
• handles struct and enum macro generation through the same syn::Result path
• updates DataBucket to 0.5.0 and data_bucket_derive to 0.3.17
• updates WorkTablesIndex to 0.0.3 and declares rkyv 0.8.17 as the minimum
• removes the unused rkyv 0.7 and other unused dependencies from the proc-macro crate
• replaces archived-row raw-pointer walking with bounds-checked offsets and unconditional rkyv validation
• returns structured DataDecodeError values for unsupported schemas and short, malformed, or misaligned archives
• documents and tests why field validation retains the complete serialized prefix for out-of-line relative pointers

Breaking API and versioning

DataBucket is bumped to 0.5.0 because this soundness correction intentionally changes public decoding contracts:

• parse_archived_row now returns Result<Vec, DataDecodeError>
• DataType::from_archived_bytes returns Result
• the legacy safe raw-pointer DataType methods are removed
• DataIterator items carry DataDecodeError

The derive crate remains 0.3.17 because its public changes are additive and its generated-code correction is compatible.

Persistence compatibility

This PR does not intentionally change the archive or page format. Generic primitive arrays deliberately retain align() == None, matching the existing u8 array implementations, and regression tests cover the old 20-byte and 32-byte behavior. rkyv guarantees format compatibility across the 0.8 release line.

Validation

• cargo build --verbose
• cargo test --verbose: 49 DataBucket tests passed
• cargo test -p data_bucket_derive
• cargo test -p data_bucket --no-default-features
• cargo clippy --all-targets --all-features -- -D warnings
• data_bucket_derive 0.3.17 publish dry-run
• WorkTable 0.9.4 all-target/all-feature downstream check with DataBucket 0.5.0 and WorkTablesIndex 0.0.3
• web3.trading-backend downstream check
• api.support.cafe workspace/all-target check with s3-sync enabled

The DataBucket package dry-run must follow publication of its exact data_bucket_derive 0.3.17 dependency. Live api.support.cafe S3 validation is intentionally deferred.

Release order after merge: data_bucket_derive 0.3.17, then data_bucket 0.5.0.

Fixes #61

[pathscale]

Disposition of the latest blocker review:

1. codegen/src/size_measure/mod.rs: the original branches already compiled because ? unwraps the enum Result to TokenStream, but the inconsistency was valid maintenance feedback. Struct generation now also returns syn::Result<TokenStream> and both branches propagate with ?.
2. enum_generator.rs: the emitted associated-type syntax was already compile-tested successfully. It now names <Enum as rkyv::Archive>::Archived explicitly and adds an explicit Archived: Sized bound. The integration derive test remains green.
3. archived-row slice anchoring: &buf[..field_end] is intentional. rkyv::access roots the field at the slice end, while the complete prefix is required for out-of-line string payloads serialized before the fixed-width row root. This is now documented inline and covered by a primitive-followed-by-long-string archive test.
4. panic-based decoding: removed. parse_archived_row and DataType::from_archived_bytes now return structured DataDecodeError values. Safe raw-pointer trait methods were removed. DataIterator propagates decode errors.
5. versioning: the public decoding change is breaking, so DataBucket is now 0.5.0. data_bucket_derive remains 0.3.17.
6. blanket DefaultSizeMeasurable: removed. default_aligned_size is now a default method on SizeMeasurable, so each concrete implementation may override it without coherence conflicts. A regression test uses a type whose Default panics and proves the override does not construct it.
7. heavy default performance: the docs state the default construction behavior and override contract. Empty String and Vec defaults do not allocate; custom expensive defaults can override directly.
8. panic tests: converted to explicit Result/error assertions for short buffers, malformed relative pointers, misalignment, and unsupported schema types.
9. dead pointer helpers: removed; repository search finds no remaining raw-pointer decode helper use.

Validation is listed in the PR body. WorkTable, web3.trading-backend, and api.support.cafe compile against the DataBucket 0.5.0 dependency chain. Live S3 validation remains deferred by request.