int8 dp4a matvec

Same 4x8 matvec tile, but weights and inputs stay packed four int8 lanes per u32 and each 4-wide dot product is one dot4I8Packed call from the packed_4x8_integer_dot_product extension. On GPUs without a native dp4a instruction this runs the extension's polyfill.

Measured in OP/s. Higher values mean better performance. Benchmark your device.