[Offload] Try loading dynamic libraries at global scope first (#227020)

Summary:
The intention of the dynamic path is to provide basic linking in cases
where the SDK is not avaialble at the user's build time. Previously this
did a distinct `dlopen` call on the library.

However, in  cases where an existing HSA implementation exists at global
scope this can cause multiple instances to be open at the same time.
Specifically this is problematic if mixing with another implementation
(Like HIP or CUDA) or trying to intercept functions (like compiler-rt).
The compiler-rt interceptors currently carry interceptors into `dlsym`
when we should instead just load these at global scope, this makes them
behave more similar to a direct link in these scenarios, which was the
intention.

The fallback remains as the current behavior, should only trigger in
cases where there is an existing HSA in the process.