v0.4.0

• New Advanced SIMD micro-kernels:
  • Matrix multiplication (MxN) of QAI8DX (dynamically quantized 8-bit integer) LHS and QSI4CX (quantized 4-bit integer) RHS with F32 output.
  • Matrix multiplication (MxN and 1xN) of BF16 LHS and RHS with F32 output.
• New SME micro-kernels:
  • SME2 F32 matrix multiplication (1xN) micro-kernels:
    • Compatible with 2VL RHS packing, for sharing one packed RHS with SME2 F32 GEMM micro-kernel.
    • Compatible with 16VL RHS packing.
  • SME F32 packing function for transposed RHS matrix.
• Enhancements to existing micro-kernels:
  • Port several quantized micro-kernels to optimized Advanced SIMD assembly.
• Register SME F32 matrix multiplication micro-kernel in the benchmark suite.
• Enable air gapped CMake builds through local third-party dependencies.