Add int8 fallback kernel for older Arm CPUs that don't support UDOT

[robertknight]

This uses a sequence of instructions to emulate UDOT using widening multiplies and pairwise adds.

In the process the helpers for instantiating GemmExecutor with different kernels were refactored into a generic trait to make it easier to exercise each of the available kernels for a given set of input and output types.

With the changes on this branch I was able to successfully run the Whisper example on a Raspberry Pi Zero 2 (Quad-core Arm Cortex A53, 512MB RAM) using the base and tiny models, at ~0.9x real time for base and ~2.1x for tiny.

[robertknight]

For future reference I note that other projects (eg. ORT, XNNPACK) use a completely different data layout for SDOT/UDOT kernels and kernels using only Neon instructions. For non-dotprod kernels they use something more like the outer product approach of the floating point kernels. Adding an emulation for the UDOT instruction was a convenient way to get something working acceptably well on these older CPUs, but suboptimal. I don't know yet how much of a difference it will make.