fix(capi): reject null and atom objects in raw vector readers

[belowzeroff]

Problem

Two C readers dereference the wrapped ray_t* without any validation, so they can crash the interpreter from pure Python:

>>> from rayforce import _rayforce_c as r
>>> r.read_vector_raw(r.RayObject())   # obj == NULL
Segmentation fault (exit 139)

read_u8_vector has the same unchecked obj->obj->type access. A second, subtler case: passing an atom to read_vector_raw passes the ray_scalar_elem_size check (it takes abs(type)), but atoms alias len with their payload, so the reader fabricates a garbage byte length for PyBytes_FromStringAndSize — an out-of-bounds read of arbitrary size.

RayObject is constructible from Python (tp_new = PyType_GenericNew, and the .pyi stub declares __init__), so nothing prevents these calls; every other reader in the binding already guards via check_type / unwrap_vec.

Fix

Bring the two stragglers in line with the rest of the binding:

• read_vector_raw: reject NULL and non-vector objects (ray_is_vec) with RuntimeError before touching type/len.
• read_u8_vector: fold the NULL check into the existing type check.

No behavior change for valid inputs — all existing callers (Vector.to_list/to_numpy/__getitem__, F32.to_python, IPC/WS send paths) pass real vectors.

Tests

test_read_vector_raw_rejects_non_vectors — bare RayObject() and an i64 atom now raise RuntimeError for both readers, and a real vector still round-trips (3 * 8 bytes). The test segfaults the interpreter on master.

Full suite: 2328 passed, 7 skipped (macOS arm64, core 2.5.0, extension rebuilt from this branch).