Implement u8 x i8 -> i32 GEMM kernel for x86_64 using AVX2 intrinsics

[robertknight]

Add packing functions for kernels that use int8 x int8 -> i32 dot product instruction sequences, and a kernel for x86_64 that uses them with AVX2 intrinsics.

This doesn't include the GEMV kernel. I will add that separately.

This is the "baseline" kernel for x64 for systems that don't support VNNI ("DL Boost") instructions.

---

Performance numbers on i5-1038NG7 (4 cores) comparing f32 matmul vs int8, excluding gemv. The int8 matmul is now faster, but other quantization overheads mean that eg. int8 ModernBERT is slightly slower than f32.

f32 x f32 -> f32
Testing kernel fma
m 512 n 512 k 512 iters 512. Duration 419.527ms (0.819ms/iter). GFLOPS 327.6
m 1024 n 1024 k 1024 iters 64. Duration 383.096ms (5.986ms/iter). GFLOPS 358.8
m 128 n 2048 k 512 iters 512. Duration 453.880ms (0.886ms/iter). GFLOPS 302.8
m 2048 n 128 k 512 iters 512. Duration 415.957ms (0.812ms/iter). GFLOPS 330.4
u8 x i8 -> i32
Testing kernel avx2-int8
m 512 n 512 k 512 iters 512. Duration 330.732ms (0.646ms/iter). GFLOPS 415.6
m 1024 n 1024 k 1024 iters 64. Duration 291.978ms (4.562ms/iter). GFLOPS 470.7
m 128 n 2048 k 512 iters 512. Duration 332.322ms (0.649ms/iter). GFLOPS 413.6
m 2048 n 128 k 512 iters 512. Duration 328.687ms (0.642ms/iter). GFLOPS 418.1