Module attention 

Scaled dot-product attention kernel.

Matches attention-kernel-v1.yaml. Attention(Q, K, V) = softmax(Q * K^T / sqrt(d_k)) * V

Each function provides one of three backends:

• fn attention_scalar(...) – Pure Rust scalar reference (ground truth)
• unsafe fn attention_avx2(...) – AVX2 SIMD implementation
• fn attention_ptx() -> &'static str – PTX assembly source string

Functions

attention_avx2^⚠

AVX2 scaled dot-product attention – delegates to scalar.

attention_ptx

PTX assembly for fused scaled dot-product attention.

attention_scalar

Scaled dot-product attention (scalar reference).