Make packing of GEMM inputs more flexible

[robertknight]

Previously the GEMM code assumed that data would be packed in the same format as the LHS / RHS input types, and with a fixed layout. To support new data types, especially int8, more flexibility will be needed.

• int8 matmuls using dot product rather than FMA instructions will use a different blocked layout

• For some architectures / data types it will make sense to expand inputs to a wider data type during packing rather than in the kernel

• For some architectures / data types the kernel can operate on an unpacked LHS matrix, if it has unit column stride. For others packing will always be required.

To enable this:

• Add Kernel methods which return descriptors specifying the size and alignment required for packing a particular A or B input.

• Modify kernel interface to use opaque [u8] slices for packing buffer contents. The kernel implementations will cast this to a slice of the type they use internally.

• Add a PackingBuffer struct which wraps a Vec<u32> buffer and provides an API for reserving space in the buffer and casting its contents to [u8] slices for the kernel and its packing methods.

The PackedAMatrix and PackedBMatrix types still have assumptions about the internal layout of packed buffers which will need to removed. That will happen in subsequent commits.

Part of #347.