fix(compiler): classify scalar pointer tiles structurally

[lucifer1004]

Summary

TileRustType::from_scalar_ptr classified scalar pointer tiles as
Kind::PrimitiveType. A scalar pointer tile (PointerTile<*mut T, {[]}>)
lowers to a structured !cuda_tile.tile type like every other tile, so the
primitive classification misroutes it in code paths that treat primitives
specially (e.g. ABI handling that assumes scalars map to plain LLVM values).

Classify it structurally as Kind::StructuredType instead.

Test

Adds a unit test asserting the structural classification. cargo test -p cutile-compiler --lib passes.

[copy-pr-bot]

This pull request requires additional validation before any workflows can run on NVIDIA's runners.

Pull request vetters can view their responsibilities here.

Contributors can view more details about this message here.

[elibol]

Thanks @lucifer1004 — you did find a real panic here, but I don't want to fix it by moving the classification.

Kind is a DSL-level thing for us, not a lowering fact. Everything lowers to !cuda_tile.tile — scalars included — so "it lowers to a tile" isn't the signal for StructuredType. Scalars and pointers are both PrimitiveType on purpose, and I'd like to keep them that way.

The actual bug is in the primitive element-type path: get_type_ident on PointerTile<*mut T, {[]}> returns the outer segment ("PointerTile"), which isn't a registered ElementType, so get_cuda_tile_element_type_primitive asserts and blows up. The structured path only survives because it walks the generic args and has a Type::Ptr arm that pulls out the pointee. So the fix belongs there — teach the primitive extractor to see through a scalar pointer tile to its pointee, the same way the structured path already does — and leave from_scalar_ptr at PrimitiveType.

Reclassifying isn't a no-op either: Kind gets dispatched in a bunch of places (binary-op operands are required to be PrimitiveType, const-value contexts too), so I'd rather not move it just to dodge the extractor gap.

If you can move the fix into the extractor and add a test that actually reproduces the panic — compile a kernel that queries the element type of a scalar pointer tile, not just an assert on the Kind — I'll happily take it. Thanks for digging it up.