[compute] Porting Q4_0 and Q8_0 weight FullyConnected compute library

[hseok-oh]

What

Porting ggml compuation library: Q4_0 FullyConnected (mulmat) into compute/cker or better place

Why

Prepare to support block quantized FullyConnected layer in CPU backend

[glistening]

We may start from single thread implementation. However, for near future:

How many threads fully connected kernel for block quantization will use?
I don't remember vividly how it works in other kernels. (e.g. ruy, eigen, ...).
As I remember, the number of threads is not controlled by core, but each kernel uses threads as it like.

_{🌳 Config.lst}

CONFIG(RUY_THREADS             , int          , "-1")
CONFIG(XNNPACK_THREADS         , int          , "-1")

Also, I don't remember what is the relationship between previous environment variable THREAD about 5 years ago.

[hseok-oh]

I think we can merge this two config to ONERT_THREADS

Usage example:

ONE/tools/stab/backend_scheduler.py

Lines 147 to 149 in c66e299

	cmd += [f"BACKENDS={';'.join(backend_list)}"]
	cmd += [f"RUY_THREADS={self.num_threads}"]
	cmd += [f"XNNPACK_THREADS={self.num_threads}"]

I'll make PR to merge config