fix(mir-importer): dispatch sync_threads via its canonical thread:: path

[haixuanTao]

Summary

Every sreg accessor arm in try_dispatch_intrinsic matches both the crate-root re-export path and the canonical module path ("cuda_device::threadIdx_x" | "cuda_device::thread::threadIdx_x"), but the barrier arm only matches "cuda_device::sync_threads". A call that resolves through the canonical cuda_device::thread::sync_threads path — e.g. a dependency crate calling the API by full path — misses intrinsic dispatch and falls through to a plain function call against a symbol no definition provides, failing module verification with "Symbol cuda_device__thread__sync_threads not found". (Found while porting a multi-crate shader codebase whose support crate calls cuda_device::thread::sync_threads() directly; mentioned in #359.)

Changes

• Add the missing "cuda_device::thread::sync_threads" alias to the barrier dispatch arm, matching the sreg arms' pattern.
• Swept the other dispatch arms against the cuda_device module tree: sync_threads was the only function with a root-path arm but no canonical-path alias.

Testing

• cargo test -p mir-importer — all green.

• Downstream validation: a support crate calling cuda_device::thread::sync_threads() by full path fails module verification without this change and compiles + runs correctly (barrier lowers to bar.sync) with it.

• cargo test passes

• New example — not applicable (one-line dispatch alias)

Checklist

• All commits signed off (git commit -s)
• SPDX headers on new source files (no new files)

🤖 Generated with Claude Code