Add missing __class_getitem__ methods to typing - #1062

Closes msgspec#1056.

Adds a `stubtest` step to the `test-typing` job so stub/runtime drift is
caught in CI instead of by hand (as in msgspec#1043, msgspec#1062). `mypy` is already
in the `test-typing` group, so no new dependency.

### Root-cause fixes (instead of allowlisting)

- **Positional-only `__text_signature__`.** `json`/`msgpack`
`encode`/`decode` and the `Encoder`/`Decoder` methods take their first
arg positional-only at runtime, but their C `__text_signature__` omitted
the `/`, so `inspect.signature` reported them as positional-or-keyword
and stubtest flagged 10 mismatches. Added the `/` in `_core.c` so
introspection matches runtime; behaviour is unchanged (keyword calls
already raised). This removes all 10 pos-only allowlist entries.
- **`StructMeta` `@disjoint_base`.** `StructMeta` is a disjoint base at
runtime; marked it `@disjoint_base` in the stub (the stub already
imports from `typing_extensions`, so no new floor of consequence).
Removes that allowlist entry.

### Other stub fixes

- `structs.pyi` was missing `__all__`.
- `json.format`s stub marked `buf` positional-only (`/`), but the
runtime accepts it as a keyword, so the `/` is dropped to match (this is
the one callable that really takes a keyword).

### Allowlist

What is left in `tests/typing/stubtest_allowlist.txt` is 10 structural
false positives stubtest cannot see through: the per-class `__init__`
synthesized in C for every `Struct`/`inspect` type, the custom `__new__`
on `Meta`/`Ext`/`UnsetType`, `Struct.__init_subclass__` (config kwargs
via `dataclass_transform`, surfaces on 3.14+), and the private `_utils`
module.

Run with `--ignore-unused-allowlist`. The `test-typing` job runs
stubtest on a single Python resolved from `python-version-file:
pyproject.toml` (currently 3.14, floats up over time), not a pinned
matrix, so a single superset allowlist with `--ignore-unused` is robust
to the runner Python moving. Verified green on 3.12, 3.14, 3.15.

---------

Co-authored-by: Siyet <Siyet@users.noreply.github.com>

Widening the runtime to accept `None` was the alternative. I rejected
it: the docstring specifies `offset : int, optional` and reserves `-1`
as the sentinel for writing at the end of the buffer, which leaves
`None` with no meaning to carry. Narrowing the stub makes the declared
type match both the C code and the documented contract.

The `int | None` traces back to the original `Optional[int]` annotation
and survived the 3.10+ syntax migration in msgspec#1041. Nothing in the tree
passes `None`.

Residual limitation: this edits the same two stub lines as open PR
msgspec#1160, which reshapes `buffer` and the return type while keeping
`offset: int | None`. Whichever lands second needs a one-line rebase.
The two changes are otherwise independent, since msgspec#1160 keeps reading
`offset` via `PyLong_AsSsize_t`.

Measurement, CPython 3.14:

- `just test-typing` clean under mypy 2.3.0, pyright 1.1.411, pyrefly
1.2.0.
- Reverting the stub hunk alone makes all three checkers fail on both
new call sites in `basic_typing_examples.py`, so the type-level test is
not vacuous.
- `pytest tests/unit`: 6119 passed, 404 skipped, identical to the
pre-change baseline. The two runtime assertions are pins on current
behavior rather than new coverage.
- `ruff format` collapses the signature onto one line once `| None` is
gone; that reformat is included.

No changelog entry, following merged stub PRs msgspec#1055, msgspec#1057, msgspec#1062,
msgspec#1065, msgspec#1074 and msgspec#1093, which leave it to the aggregated "type stub
improvements" line at release.

Co-authored-by: Dylan Pulver <dylanpulver@users.noreply.github.com>