Annier.dq finetune

[Annieren]

Overview

1. Refactor the previous the mmvdq design, make it possible to have a parallel/mirror path of mmvq.
2. Add moe feature in mmvdq design
3. Add some test cases in test-backend-ops.cpp, also add GGML_TEST_FLUSH_CACHE env (Default false) to enable flush mall for decoding kernel level test.

Requirements

• I have read and agree with the contributing guidelines
• AI usage disclosure:
  Yes, for test and documentation.

Test Result:

1. We can see moe decoding performance improvement:

model	dqoff tg med	dqon tg med	tg dqon/dqoff%	hangs
Cosmos-Reason2-8B-Q8_0	26.68	26.69	+0.0	0
Qwen3.5-9B-Q4_0	40.07	40.30	+0.6	0
Qwen3-30B-A3B-Q4_K_M	80.40	82.98	+3.2	0
Qwen3.5-9B-Q4_K_M	37.55	37.73	+0.5	0
gemma-4-12b-it-Q4_K_M	28.03	28.41	+1.4	0
Qwen3.6-35B-A3B-UD-Q4_K_XL	58.12	59.05	+1.6	0
Qwen3.6-35B-A3B-UD-Q5_K_XL	56.55	57.00	+0.8	0
Qwen3.5-35B-A3B-Q4_K_M	59.28	60.15	+1.5	0
gemma-4-26B-A4B-it-UD-Q4_K_M	50.25	50.86	+1.2	0
Qwen3-8B-Q4_K_M	43.88	43.52	-0.8	0
Meta-Llama-3.1-8B-Instruct-Q4_K_M	44.77	44.41	-0.8	0

2. PPL has slightly improvement or equal with fp32 activation path with current implementation so far. No regression. Full ppl data will be posted once all the kernels are ported.