[MAHOUT-1414][QDP] Expose native cuda_available() and gate GPU tests on it

[ryankert01]

Closes #1414.

Why

#1321 lets _qdp build and import without the CUDA toolkit (stub CUDA Runtime symbols). That breaks the long-standing assumption that "the _qdp extension imports" ⟺ "a GPU is usable", which the test suite relied on to gate GPU tests. #1321 patched the immediate breakage by skipping @pytest.mark.gpu tests on torch.cuda.is_available(), but that is a proxy: on this feature's headline scenario — a GPU host with PyTorch but no toolkit — _qdp is a stub build, yet torch.cuda.is_available() is still True, so the GPU tests would run against the stub and abort.

This PR makes the engine report the truth.

What

• qdp-core: add cuda_runtime_available() — queries cudaGetDeviceCount and returns true only when a device is actually present. In a qdp_no_cuda (stub) build it resolves to the existing 999-sentinel stub, so it returns false with no cfg branching or dead code. Re-exported from the crate root.
• _qdp: expose it as _qdp.cuda_available().
• qumat_qdp: add is_cuda_available() (mirrors the existing is_triton_amd_available()) as the single Python source of truth.
• testing/conftest.py: gate the @pytest.mark.gpu auto-skip on the native signal, falling back to torch only if the helper is absent.
• test_fallback.py: coverage that also runs on a stub build, guarding that querying availability returns a bool without aborting.

Test plan

Verified locally on 4× NVIDIA GPUs (CUDA 12.8), real build:

• _qdp.cuda_available() / qumat_qdp.is_cuda_available() → True with a GPU, False under CUDA_VISIBLE_DEVICES="".
• testing/qdp + testing/qdp_python with CUDA hidden: 130 passed, 156 skipped, 0 failed (no aborts).
• The 10 tests that crashed in fix(qdp-core): link cudart locally and stub FFI when toolkit is absent  #1321's CI run: pass on GPU, skip when hidden.
• qdp-core builds in both real and QDP_NO_CUDA=1 stub modes; full Rust suite, clippy --all-features -D warnings, ruff, and ty all clean.

Out of scope (left in #1414)

Making the stub path raise a clean RuntimeError instead of aborting for direct (non-test) callers, and collapsing the ~65 redundant inline torch.cuda.is_available() checks now that the conftest gate is accurate.

[rich7420]

@ryankert01 thanks for the patch, Nice one
locally tested
the no-cfg-branch stub resolve is clean, and gating on the native signal instead of torch is the right fix. Verified both crates build in stub mode.

Two tiny nits, both optional:

1. Three names for one concept (cuda_runtime_available → cuda_available → is_cuda_available). Each follows local convention so it's fine, just a little to track.(it's not that important)

2. is_cuda_available() does a fresh FFI cudaGetDeviceCount every call, and conftest calls it per gpu test. Cheap so not a perf issue, but the result can't change within a run — an lru_cache (or OnceLock<bool>) would make it a one-time query.

[viiccwen]

LGTM!