metal : dequantize q8_0 using packed types

[ggerganov]

Overview

Extract the packed_char4 pattern from @tarruda's #25556. This is a free win for -ctk q8_0 -ctv q8_0 on Metal with large contexts and it's a simple change.

Additional information

./scripts/compare-commits.sh master gg/metal-dequantize-q8_0-packed llama-bench -hf ggml-org/Qwen3.6-35B-A3B-GGUF:Q8_0 -ub 2048 -d 0,8192,16384,32768,65536 -n 32 -r 5 -fa 1 -ctk q8_0 -ctv q8_0

| CPU           | Model                  | Test         |   t/s master |   t/s PR  |   Speedup |
|:--------------|:-----------------------|:-------------|-------------:|----------:|----------:|
|Apple M2 Ultra | qwen35moe 35B.A3B Q8_0 | pp512        |      1976.83 |   1989.83 |      1.01 |
|Apple M2 Ultra | qwen35moe 35B.A3B Q8_0 | pp512@d8192  |      1337.86 |   1338.22 |      1.00 |
|Apple M2 Ultra | qwen35moe 35B.A3B Q8_0 | pp512@d16384 |      1004.65 |   1004.76 |      1.00 |
|Apple M2 Ultra | qwen35moe 35B.A3B Q8_0 | pp512@d32768 |       672.73 |    671.86 |      1.00 |
|Apple M2 Ultra | qwen35moe 35B.A3B Q8_0 | pp512@d65536 |       399.99 |    399.37 |      1.00 |
|Apple M2 Ultra | qwen35moe 35B.A3B Q8_0 | tg32         |        74.07 |     75.17 |      1.01 |
|Apple M2 Ultra | qwen35moe 35B.A3B Q8_0 | tg32@d8192   |        58.59 |     61.40 |      1.05 |
|Apple M2 Ultra | qwen35moe 35B.A3B Q8_0 | tg32@d16384  |        46.21 |     50.30 |      1.09 |
|Apple M2 Ultra | qwen35moe 35B.A3B Q8_0 | tg32@d32768  |        32.57 |     36.96 |      1.13 |
|Apple M2 Ultra | qwen35moe 35B.A3B Q8_0 | tg32@d65536  |        20.42 |     23.98 |      1.17 |

Requirements

• I have read and agree with the contributing guidelines
• AI usage disclosure: YES - see referenced PR