[Core ML] Avoid unnecessary portable-kernel linkage from coremldelegate

Problem

The Core ML CMake target links portable_ops_lib and portable_kernels privately when the standalone runner option is enabled. The standalone Xcode runner already links those archives itself, while other Core ML consumers should not pull in a second operator registry.

A proposed switch from link edges to build dependencies surfaced another requirement: add_dependencies fails when portable operators are disabled unless both targets are guarded. This was found during the source-build validation for #22620 and is intentionally excluded from that documentation-only PR.

Proposed change

Use build-order dependencies instead of private linkage where appropriate, guarded with TARGET portable_ops_lib and TARGET portable_kernels.

Test plan

• Configure the macOS preset with Core ML enabled and portable operators disabled.
• Build the Core ML runner configuration with portable operators enabled.
• Run Core ML model and operator CI.

cc @larryliu0820 @GregoryComer @kimishpatel @YifanShenSZ @cymbalrush @metascroy