[Annieren]

Overview

1. Add Q4_0 direct fp32 activation kernel support.
2. Add Q4_0 in test-backend-ops: correctness and performance
3. Update Q8_0 test-backend-ops correctness support

Additional information

1. Q4_0 model decoding test gains on top of add q8_0 data type direct fp32 activation support #142.

==== decode (tg128) ====

model	base	new	Δ (B/A)
gemma-4-E2B-it-Q4_0	104.87	109.55	+4.5%
gemma-4-E4B-it-Q4_0	58.67	60.66	+3.4%
Qwen3.5-2B-Q4_0	131.27	134.01	+2.1%
Qwen3.5-4B-Q4_0	66.53	67.75	+1.8%
Qwen3.5-9B-Q4_0	40.19	40.38	+0.5%
Qwen3.8-27B-Q4_0	13.39	13.40	+0.0%
mean Δ = +2.1% (n=6)

2. No impact on prefill
3. Passed PPL test. very slightly improvement over mmvq path

Requirements

• I have read and agree with the contributing guidelines
• AI usage disclosure:

Yes. AI is used for document and test.