auto-round

b56f301b - perf(ark): 2D block store epilogue + single-pass activation quantizer for W4A8 MoE prefill

perf(ark): 2D block store epilogue + single-pass activation quantizer for W4A8 MoE prefill Replaces the W4A8 prefill epilogue's 64 scalar 32-byte stores per sub-group fragment with the hardware 2D block store (`make_block_2d_copy_D` + `copy(copy_d, tCrD, tCgC)`, the sequence already compiled in `sycl_tla_dense_gemm.hpp` for the same accumulator/output widths), and removes the activation quantizer's second read of every row by holding it in registers between the absmax and the quantize pass. Also corrects the harness roofline, which counted only weight bytes and so understated the bandwidth a shape needs to hit 100 TFLOPS by up to 2.2x. Co-authored-by: a32543254 <53296245+a32543254@users.noreply.github.com>