CANN: Optimize MUL_MAT_ID

[hipudding]

./bin/test-backend-ops test -b CANN0 -o MUL_MAT_ID
Testing 5 devices

Backend 1/5: CANN0
Device description: Ascend910B4
Device memory: 30196 MB (29803 MB free)

new_pool_for_device: device 0 use vmm pool
11837/11837 tests passed
Backend CANN0: OK
Backend 2/5: CANN1
Skipping
5/5 backends passed
OK

Performance result for qwen3:30b-a3b-fp16

Before

llama_perf_context_print:        load time =   32008.05 ms
llama_perf_context_print: prompt eval time =     465.31 ms /    16 tokens (   29.08 ms per token,    34.39 tokens per second)
llama_perf_context_print:        eval time =    3771.73 ms /    12 runs   (  314.31 ms per token,     3.18 tokens per second)
llama_perf_context_print:       total time =    4766.22 ms /    28 tokens
llama_perf_context_print:    graphs reused =         12

After

llama_perf_sampler_print:    sampling time =      54.81 ms /   224 runs   (    0.24 ms per token,  4086.62 tokens per second)
llama_perf_context_print:        load time =   30221.19 ms
llama_perf_context_print: prompt eval time =     298.75 ms /    16 tokens (   18.67 ms per token,    53.56 tokens per second)
llama_perf_context_print:        eval time =    5157.83 ms /   207 runs   (   24.92 ms per token,    40.13 tokens per second)
llama_perf_context_print:       total time =    6018.31 ms /   223 tokens
llama_perf_context_print:    graphs reused =        207

Make sure to read the contributing guidelines before submitting a PR