test: ffi_tests::test_memory_functions fails deterministically on CUDA (wired limit is Metal-only) #741

Context

src/lib/mlxcel-core/src/ffi_tests.rs test_memory_functions asserts set_wired_limit(1 GiB) and then get_wired_limit() > 0. The wired-memory limit is an Apple Silicon / Metal concept; on the CUDA backend (GB10) get_wired_limit() returns 0, so the assert at ffi_tests.rs:1756 fails deterministically:

cargo test --release -p mlxcel-core --features cuda --lib test_memory_functions -- --test-threads=1

fails in isolation (0.31 s, assertion failed: limit > 0).

Discovered on 2026-07-11 while running the mlxcel-core ffi regression suite for PR #740 (92/93 passed, this was the only failure). The test predates the #725 work (introduced in commit 0718bce, "Separate mlxcel-core FFI wrappers") and is unrelated to the quantized-matmul changes. It appears the mlxcel-core lib test suite is not part of the root-package CI/test runs on CUDA, which is why it went unnoticed.

What's needed

• Gate the wired-limit assertions to the Metal backend (e.g. cfg(feature = "metal") or a runtime backend check), or make the CUDA expectation explicit (assert the call is a no-op returning 0 on CUDA).
• Consider whether the mlxcel-core lib tests should run in CI for the cuda feature so backend-conditional tests get exercised.

Acceptance Criteria

• test_memory_functions passes on both Metal and CUDA hosts
• The backend-conditional expectation is documented in the test comment
• A note on whether mlxcel-core lib tests are included in the CUDA test entry point