Working INT8/INT4 weight-only quantization

Decode is bandwidth-bound: `benchmarks/roofline.py` measures 160 GB/s achievable against 942 MiB of weights, a 6.22ms/step floor. Halving weight bytes should roughly halve decode time. This is the largest single win available. **Prior attempt (failed, documented in BENCHMARKS.md §10):** torchao `Int8WeightOnlyConfig` made it 2.5x *slower* (11.5 → 29.1 ms/step) and produced garbage output. Eager mode dequantizes to fp16 before the matmul, which adds traffic rather than removing it, and the tensor subclasses do not survive our hand-written forward pass. **What is actually needed:** a fused dequant-matmul (Marlin-style) so weights are read as int8/int4 and expanded inside the kernel. **Done when:** measured ms/step improves at B=1..16 and output quality is checked, not assumed.