Loop unrolling hints

[tetsuo-cpp]

Summary

Add pragma-style annotations for loop unrolling to enable compiler-driven performance optimization.

Motivation

Inner loops in GPU kernels (especially the K-dimension loop in matmul) benefit significantly from unrolling. The compiler can sometimes unroll automatically, but explicit hints give more control.

Design considerations

Possible syntax approaches:

\ Attribute-style before loop
#UNROLL 4
DO ... LOOP

\ Or a dedicated word
4 UNROLL-DO ... LOOP

Implementation notes

• Could map to LLVM loop metadata (llvm.loop.unroll.count)
• Or use MLIR's affine.for with unroll attributes
• Need to decide if this is a parser-level annotation or a runtime construct
• Should support both full unroll and partial unroll (by factor)

Priority

Nice to have — performance optimization for inner loops.