UCP/RNDV: Add opt-in intra-node CUDA staging policy

Allow users to prefer the copy-to-attached host-staged path for intra-node CUDA-to-CUDA tag rendezvous while leaving explicit schemes and unrelated traffic unchanged. Force only the staged receiver envelope and the sender-side rndv/rkey_ptr/mtype child when that child initializes, with normal UCX fallback otherwise.

Mark internal RMA and AM rendezvous selections with local-only provenance and preserve it across pipeline fragment selection, so RMA PUT/GET reuse of RNDV_RECV and Active Message rendezvous remain outside the tag-only policy without changing the wire format.

Scope synthetic receiver-estimation rkey identity to matching forced intra-node CUDA modeling, so disabled and unrelated flows reuse the legacy rkey configuration.