feat: add shared memory support via \! shared header directive

[tetsuo-cpp]

Closes #4

Summary

• Add \! shared <name> i64[<N>] header directive for declaring GPU shared (workgroup) memory
• ForthToMLIR emits tagged memref.alloca at kernel entry; using the shared name pushes its base address onto the stack
• ForthToGPU converts tagged allocas to gpu.func workgroup attributions (memref<Nxi64, #gpu.address_space<workgroup>>)

Test plan

• Translation test: shared alloca emission with forth.shared_name attr and pointer push sequence
• Translation error test: shared name rejected inside word definitions
• Translation error test: param/shared cross-namespace name collision
• ForthToGPU conversion test: alloca replaced by workgroup attribution
• End-to-end pipeline test: full pipeline produces gpu.binary, intermediate check for workgroup attribution
• All 72 tests pass

[tetsuo-cpp]

Code Review

Clean, well-structured PR. The design of using memref.alloca + marker attribute as IR, then converting to workgroup attributions in ForthToGPU, is a nice approach that avoids a new dialect op.

Issues

1. Stale TODO — ForthToMLIR.h:29

/// TODO: Not yet consumed — scaffolding for shared memory support.
struct SharedDecl {

SharedDecl is now fully consumed — this TODO should be removed.

2. Comment says "three transformations" but lists two — ForthToGPU.cpp:161-163

// block, with three transformations:
// - func.return → gpu.return
// - shared memref.alloca → gpu.func workgroup attribution

Should say "two transformations" (or add the implicit third: everything else gets cloned).

3. Scalar shared declarations silently accepted
The shared parsing path (\! shared X i64, no array suffix) is accepted since it shares code with \! param. A scalar shared variable is unusual — consider rejecting it or documenting it. CLAUDE.md currently only documents the array form.

4. No test for param/shared cross-namespace name collision
The duplicate detection correctly catches e.g. \! param A i64[4] followed by \! shared A i64[8], but no test exercises this path. Only param-param collision is tested (header-duplicate-param-error.forth).

What looks good

• Clean separation: ForthToMLIR handles parsing/emission, ForthToGPU handles address space conversion
• IRMapping approach in ForthToGPU elegantly maps alloca → workgroup BlockArgument so all downstream uses automatically see the workgroup-typed memref
• Good error message for shared refs inside word definitions
• Solid test coverage across translation, conversion, and pipeline layers