Mature the pure-Rust GPU graph executor behind capability reporting

[GeneralPawz]

Roadmap item, Parked: performance and execution providers.

Parked deliberately. The GPU crate is currently a seam
(crates/execution/gpu, ~773 lines), not a mature executor. Saying so plainly
is more useful than implying capability.

Why parked

Same ordering principle as #16: capability before speed.

Unblocked when

The exact path (v0.2) is real, so there is something worth accelerating.

Done when

• Capability reporting states what the GPU path can actually execute
• Scalar differential checks gate every GPU result
• No silent CPU fallback that hides an unsupported operation
• Same-harness evidence before it is enabled by default