optimize for ppc64le using VSX intrinsics

[penghongbo]

I would like to contribute to the ggml in improving the performance on ppc64le. Here is the updated code by using the VSX intrinsics. Please review. Thank you.

[penghongbo]

@ggerganov would you please review this PR or assign reviewers? I am working for IBM and we would like to have this optimization on ppc64le. Thank you.

[penghongbo]

@ggerganov is this your preferred place to make PR for enhancement or optimization. Or shall I make PR in llama.cpp and then you sync to this repository?

[ggerganov]

Hey, thanks and sorry for the delay. This implementation fits in the existing pattern and seems all new code is behind the __POWER9_VECTOR__ define which is great.

I don't have the hardware to test this - can you share some sample numbers for the speed-up that you observe?

[penghongbo]

Here is the speedup of vec_dot_q float32 throughput numbers compared to current master branch on RHEL9.2 (Power10 machine) by the test-quantizer-perf -i 10000. The code was compiled by gcc-12.2.1.

Type	q4_0	q4_1	q5_0	q5_1	q8_0	q2_K	q3_K	q4_K	q5_K	q6_K	iq3_xxs	iq4_nl	iq3_s	iq2_s	iq4_xs
speedup	4.99	3.39	4.27	3.17	1.37	7.70	4.56	5.26	4.78	5.14	3.57	5.97	3.61	4.91	6.99

The float32 throughput of q8_0 in function quantize_row_q and quantize_row_q_dot also get a speedup as around 4.81.

I also tried to run the test-quantizer-fns and verified the code in llama.cpp to verify the functionality. All passed in the test env as mentioned above.

Thanks for your review and approval.