Fix ModuleProxy semantic dispatch for Ascend

Restore the libshmem_device proxy API surface by removing explicit _semantic parameters from the dispatch stubs, and teach ModuleProxy.dispatch to advertise _semantic in its generated builtin wrapper signature.

This is needed because the Ascend backend uses triton-ascend, whose Triton 3.5-style code generator only injects _semantic when it appears in the callable signature. The previous proxy wrapper accepted **kwargs, which was enough for Triton 3.4 backends where _semantic was always injected, but its wrapped stub signature did not expose _semantic to signature introspection.

Adding _semantic at the ModuleProxy layer keeps the shared proxy API clean while allowing Triton 3.5 to pass the compiler semantic context through to Ascend extern implementations. Forward _semantic only when the selected backend method accepts it, preserving compatibility with CUDA, HIP, and MACA implementations that may use different internal conventions.

Author: janssen-v