fix(hnsw): reject build dims whose vector count overflows size_t

[belowzeroff]

What & why

ray_hnsw_build sized the copied vector block as n_nodes * dim * sizeof(float)
with no overflow check (src/store/hnsw.c). Dimensions whose product wraps
size_t under-allocate the copy, while the following memcpy — and every later
distance read (vectors + id*dim) during graph construction — runs past the
buffer.

The per-layer neighbor arrays in the loader already carry a divide-based
overflow guard; the build path had none. This adds the same guard at the top of
ray_hnsw_build, rejecting overflowing dimensions before any allocation:

if ((uint64_t)n_nodes > SIZE_MAX / sizeof(float) / (uint64_t)dim) return NULL;

n_nodes and dim are already validated > 0 at that point, so the divide is
safe.

Scope / severity: this hardens the public C API boundary. The in-tree
(hnsw-build ...) builtin sizes its vectors from an in-memory list and cannot
reach the overflow, so for the shipped language surface this is defense-in-depth
rather than a remotely reachable bug. It is the build-path companion to the
loader's untrusted-file vectors guard (submitted separately).

Test

embedding/hnsw_build_overflow_rejected drives an overflowing n_nodes/dim
pair (chosen so the byte count wraps to 16) and asserts the build is refused.
Verified as a genuine regression test: with the guard removed, the debug ASan
build faults at the copy —

SUMMARY: AddressSanitizer: stack-buffer-overflow hnsw.c:494 in ray_hnsw_build

With the guard in place the call returns NULL and the source is never read.

Full suite under the debug ASan + UBSan build: 3632 of 3633 passed (1 skipped, 0 failed).

Checklist

• PR targets dev (not master)
• Commits follow Conventional Commits (fix:)
• make builds cleanly (no new warnings)
• make test passes; tests added/updated for behaviour changes