Port of base.py quantized_scaled_dot_product_attention (stock, non-tiled): scores and output via quantized_matmul against the quantized KV triples; GQA via a 5-d reshape. Exported for the fused-vs-unfused parity tests.
Port of base.py quantized_scaled_dot_product_attention (stock, non-tiled): scores and output via quantized_matmul against the quantized KV triples; GQA via a 5-d reshape. Exported for the fused-vs-unfused parity tests.