Shared-memory declarations for reduced-width element types

[tetsuo-cpp]

Summary

Allow \! shared buffers declared as i8/i16/i32/f16/bf16/f32 (not only i64/f64), matching the existing S-prefixed narrow load/store words.

Motivation

Access words already exist (SI8@, SHF!, SF32@, etc.), but shared declarations are effectively i64/f64-oriented. Tiled kernels often want f32 or f16 tiles in shared memory with correct sizing (bytes vs cells) and allocation footprint.

Proposed syntax

\! shared TILE f32[256]
\! shared MASK i8[128]

Lowering:

• Workgroup attribution sized by sizeof(elem) * N
• Base address usable with the matching S* access words (or element-scaled indexing from the typed-params issue)

Acceptance criteria

• Parser accepts reduced-width shared declarations
• ForthToGPU / workgroup memory size is correct in bytes
• LIT tests for each major width class (int + float)
• GPU test storing/loading f32 shared tile
• Docs updated

Related

• Shared memory (Shared memory support for GPU scratchpad #4 closed) and S@/S! split (Proposal: split global vs shared memory words (@/! and S@/S!) #36 closed)
• Reduced-width ops (Add load/store words for reduced-width memory types #52 closed)
• Element-typed parameters (separate issue)