Replace private GEP inbounds metadata with upstream Pliron no-wrap flags #1231

CUDA_OXIDE currently preserves the distinction between ordinary and wrapping pointer offsets using private LLVM-dialect metadata.

This was introduced in #612 to fix wrapping pointer-offset lowering: ordinary pointer offsets may carry the LLVM inbounds contract, while wrapping raw-pointer offsets must not be emitted as getelementptr inbounds.

At the moment, that distinction is represented in two layers:

• MIR keeps mir_ptr_offset_inbounds, which models the Rust-level semantic distinction.
• The LLVM dialect carries a private cuda_oxide_gep_inbounds marker consumed by llvm-export.

Pliron now has first-class support for LLVM GEP no-wrap flags through pliron-org/pliron#244:

• inbounds
• nusw
• nuw
• canonical inbounds => nusw semantics
• GetElementPtrOp::new_with_no_wrap_flags
• preservation across LLVM IR conversion

CUDA_OXIDE should adopt that upstream representation and remove the private LLVM-level shim.

Scope

• Keep the MIR-level mir_ptr_offset_inbounds marker because it represents Rust source semantics.
• Lower ordinary pointer offsets to GEPs carrying GepNoWrapFlags::INBOUNDS.
• Lower wrapping pointer offsets to GEPs with no no-wrap flags.
• Replace existing production GEP sites that previously relied on CUDA_OXIDE's implicit inbounds behavior with explicit upstream flags.
• Keep dynamic shared-memory byte-offset GEPs unflagged: the [0 x i8] declaration does not establish LLVM's allocated-object inbounds contract for non-zero runtime offsets.
• Teach llvm-export to emit Pliron's INBOUNDS, NUSW, and NUW flags.
• Remove cuda_oxide_gep_inbounds and its associated helpers.
• Add regression coverage for the supported flag combinations and pointer-offset semantics.

Acceptance criteria

• No remaining uses of the private cuda_oxide_gep_inbounds representation.
• Ordinary pointer offsets continue to export as getelementptr inbounds.
• Wrapping pointer offsets export as plain getelementptr.
• Dynamic shared-memory addressing remains plain GEP.
• plain, nusw, nuw, inbounds, and inbounds nuw export correctly.
• Workspace tests and Clippy pass.
• Full compile-only smoketest suite passes.
• ptr_offset_from and dynamic_smem pass on real GPU execution.