ggml-cpu: add RVV implementation for q1_0 x q8_0 vec dot

[velonica0]

Overview

This PR adds an RVV-specific implementation of ggml_vec_dot_q1_0_q8_0 for RISC-V.

This PR is a follow-up to #21273 and #21636, and my changes here are based on the work from those PRs. Thanks to the authors for laying the groundwork.

Benchmark results for Bonsai 1.7B (Due to runtime constraints on the K1, the benchmark was limited to 64 token)

threads	test	generic tok/s	RVV tok/s	speedup
4	pp64	0.561362 +- 0.000273	2.462494 +- 0.007606	4.39x
4	tg64	0.442725 +- 0.000522	1.872843 +- 0.005062	4.23x
8	pp64	0.880189 +- 0.002553	3.418164 +- 0.015693	3.88x
8	tg64	0.367250 +- 0.005271	0.796739 +- 0.005334	2.17x

Perplexity summary style table:

Metric	RVV (K1)
Same top p	99.294 +- 0.235 %
Mean KLD	0.000222 +- 0.000009
Maximum KLD	0.004316
99.9% KLD	0.004056
99.0% KLD	0.001395
Median KLD	0.000137
1.0% KLD	-0.000012
Minimum KLD	-0.000051
Mean Delta p	-0.007 +- 0.010 %
Maximum Delta p	2.477 %
99.9% Delta p	2.177 %
99.0% Delta p	1.180 %
95.0% Delta p	0.460 %
Median Delta p	-0.000 %
5.0% Delta p	-0.546 %
1.0% Delta p	-1.088 %
0.1% Delta p	-1.604 %
Minimum Delta p	-1.646 %
RMS Delta p	0.346 +- 0.017 %

Additional information

Requirements

• I have read and agree with the contributing guidelines
• AI usage disclosure:
  Use AI to help me with testing and data collection.