fix: worktable_version! does not compile for an unsized primary key

[pathscale]

The bug

worktable_version! fails to compile for any table whose primary key is not
fixed-size. A String key is the common case, and it is the case that matters:
a table keyed on a String cannot have a schema migration written against it at
all. Changing only the key type takes the same macro invocation from 0 errors to 4.

worktable_version!(
    name: Thing,
    columns: { id: String primary_key, name: String },
    indexes: { name_idx: name },
);

error[E0061]: this method takes 1 argument but 0 arguments were supplied
error[E0308]: mismatched types: expected `ThingWorkTable`, found future
error[E0061]: this struct takes 1 argument but 2 arguments were supplied
error[E0609]: no field `1` on type `&ThingWorkTable`

Cause

read_only and pk_unsized are independent flags on the PersistTable derive.
read_only selects the read-only shape: no persistence engine or task, a sync
into_worktable(), and a 1-tuple table struct. pk_unsized separately selects
the unsized primary index.

The read-only table generator emitted them as if they were mutually exclusive
(codegen/src/generators/read_only/table/mod.rs), so an unsized key silently
dropped read_only and generated the full persist shape against a read-only
table struct. load() then called an into_worktable() that had been generated
as async and returning Self(table, PersistenceTask), against a struct with one
field, which is exactly the four errors above.

The generator type is not involved. pk_gen_state() emitting
self.0.pk_gen.get_state() unconditionally is fine, since the state impl exists
for every generator including GeneratorType::None.

Fix

Emit #[table(read_only, pk_unsized)]. The attribute parser already accepts both
via parse_nested_meta, and every downstream read_only branch already composes
with pk_unsized: into_worktable's read-only arm uses the shared
primary_index_init, and the engine and task types return empty regardless. The
sized path is byte-for-byte unchanged.

Tests

Two, both verified to fail before the fix and pass after:

• tests/worktable_version/string_primary_key.rs writes two rows through a
  persisted String-keyed table, reloads them through worktable_version!, and
  asserts on both a full scan and a lookup by string key. Without the fix it
  reproduces the four errors verbatim.
• codegen/src/worktable_version/mod.rs gains a unit test asserting the emitted
  attribute keeps read_only for an unsized key.

The path was never covered because tests/worktable_version/basic.rs, the only
user of the macro, uses u64 primary_key autoincrement.

Release

Bumps worktable and worktable_codegen to 0.9.3 with the exact pin, matching
the 0.9.2 release commit. CHANGELOG.md is untouched, consistent with the 0.9.x
releases. Publishing is left for a human to run from merged master per
AGENTS.md.

Verification

cargo test 320 passed / 0 failed / 4 ignored (the 4 ignores are pre-existing and
annotated as such on master), 115 unit and 45 codegen tests pass,
cargo clippy --all-targets clean, cargo fmt --check clean.