Add GPU-guarded numerical equivalence tests for native attention kernels

[adohe]

Finding

The native backend's selling point is logprob consistency, yet no test asserts areno_varlen_causal_attention / areno_paged_causal_attention_decode match flash-attn or SDPA numerically. The only test on the used path stubs fake_varlen and checks shapes, not math.

Acceptance

• Add a GPU-guarded equivalence test (skips when no GPU) comparing used kernels against a reference.
• CPU CI remains unaffected.