Try to use quantized ggml_mul_mat in attention layer

[ggerganov]

The following 2 matrix multiplication calls sill remain in FP16 precission:

• https://github.com/ggerganov/llama.cpp/blob/d40fded93e1a533e969768e1e335c15c61c296ce/llama.cpp#L1135-L1137
• https://github.com/ggerganov/llama.cpp/blob/d40fded93e1a533e969768e1e335c15c61c296ce/llama.cpp#L1158-L1160

Was wondering, if we quantize those on-the-fly would there be any benefit.
The quantization can be done with an extra ggml_cpy() call, before the ggml_mul_mat() call.

See if this speeds up the computation and how it affects perplexity

[ggerganov]

Performance stats on 7B indicate that the F16 matrix multiplication account for just 2% of the total processing time.
The PoC in #1103 does not show any significant performance gains, so closing this for now