opencl: add optimized q4_1 mm kernel for adreno

[shaofeiqi]

This PR adds optimized OpenCL kernels for Q4_1 GEMM and GEMV operations on Adreno GPUs.

[lhez]

On X Elite,

Before,

Qwen3-0.6B-Q4_1

common_perf_print: prompt eval time =     532.78 ms /   235 tokens (    2.27 ms per token,   441.08 tokens per second)
common_perf_print:        eval time =    5592.80 ms /   256 runs   (   21.85 ms per token,    45.77 tokens per second)

Qwen3-4B-Q4_1

common_perf_print: prompt eval time =    3652.33 ms /   235 tokens (   15.54 ms per token,    64.34 tokens per second)
common_perf_print:        eval time =   23390.10 ms /   256 runs   (   91.37 ms per token,    10.94 tokens per second)

Llama-3.2-3B-Instruct-Q4_1

common_perf_print: prompt eval time =    2826.35 ms /   236 tokens (   11.98 ms per token,    83.50 tokens per second)
common_perf_print:        eval time =   19356.50 ms /   256 runs   (   75.61 ms per token,    13.23 tokens per second)

After,

Qwen3-0.6B-Q4_1

common_perf_print: prompt eval time =     213.15 ms /   235 tokens (    0.91 ms per token,  1102.53 tokens per second)
common_perf_print:        eval time =    5573.82 ms /   256 runs   (   21.77 ms per token,    45.93 tokens per second)

Qwen3-4B-Q4_1

common_perf_print: prompt eval time =    1099.39 ms /   235 tokens (    4.68 ms per token,   213.75 tokens per second)
common_perf_print:        eval time =   21173.46 ms /   256 runs   (   82.71 ms per token,    12.09 tokens per second)

Llama-3.2-3B-Instruct-Q4_1

common_perf_print: prompt eval time =     776.87 ms /   236 tokens (    3.29 ms per token,   303.78 tokens per second)
common_perf_print:        eval time =   15153.24 ms /   256 runs   (   59.19 ms per token,    16.89 tokens per second)