feat: add reduced-width memory load/store words

[tetsuo-cpp]

Summary

• Add 24 new ops for reduced-width memory access: 6 types (i8, i16, i32, f16, bf16, f32) x load/store x global/shared
• Integer loads sign-extend to i64; integer stores truncate from i64
• Float loads extend to f64 then bitcast to i64; float stores bitcast to f64 then truncate
• Generalize MemoryLoadOpConversion/MemoryStoreOpConversion templates with MemType enum

Closes #52

Test plan

• Translation test: all 24 Forth words produce correct forth.* ops
• Conversion test: verify extsi/trunci for integers, extf/truncf+bitcast for floats, ptr<3> for shared
• Pipeline test: end-to-end compilation to gpu.binary
• All 109 tests pass

[tetsuo-cpp]

Code Review: Add reduced-width memory load/store words

Overview

Adds 24 new Forth operations for reduced-width memory access across 6 types (i8, i16, i32, f16, bf16, f32) × load/store × global/shared. The implementation cleanly generalizes the existing MemoryLoadOpConversion/MemoryStoreOpConversion templates by replacing the bool IsFloat parameter with a MemType enum, then instantiates all 24 variants as type aliases. Well-structured change with good test coverage at all three layers.

Issues

Summary/description mismatch in ODS definitions (bug)

All six integer load op summaries say "zero-extend" but the descriptions say "sign-extends" and the implementation uses arith::ExtSIOp (sign-extend). The summaries need to say "sign-extend" to match:

• Forth_LoadI8Op: "Load i8 value from memory, zero-extend to i64" → should be "sign-extend"
• Forth_SharedLoadI8Op: same
• Forth_LoadI16Op: same
• Forth_SharedLoadI16Op: same
• Forth_LoadI32Op: same
• Forth_SharedLoadI32Op: same

The descriptions within the [{ }] blocks are correct — only the let summary strings are wrong.

Suggestions

• Conversion test coverage: test/Conversion/ForthToMemRef/reduced-width-memory.mlir only tests shared memory for i8. Since the template is identical, this is low-risk, but adding even one shared-memory check for a float type (e.g. shared_load_f16 with !llvm.ptr<3>) would catch any future AddressSpace regression without much cost.

• Pipeline test: Similarly only tests a subset (I8, I32, HF, F32) — covers both reduced-int and reduced-float code paths, which is sufficient. Could optionally add I16 or BF16 to be thorough, but not strictly necessary.

Positive Notes

• Template generalization is clean: The MemType enum + isReducedInt/isReducedFloat constexpr helpers + getMemElemType is a nice pattern. The constexpr if branching covers all cases exhaustively.
• Follows project conventions: CamelCase naming, stack effect comments, consistent assembly format via Forth_StackOpBase.
• CLAUDE.md updated: Both the supported words list and the new "Reduced-Width Memory" section are correct and consistent with the implementation.

Verdict

One documentation bug (summary says zero-extend, should say sign-extend) that should be fixed before merge. Otherwise the code is correct and well-organized.