tests: support multi-op perf groups in test-backend-ops

[zzzzwc]

Overview

• Refactor eval_perf() to support fused multi-op benchmarking: perf_group_size() declares how many trailing nodes form the effective operation; only those are duplicated per run, setup nodes run once.
• Add --perf-duration and --threads CLI options.
• Add test_rms_norm_mul perf test cases for broadcast and non-broadcast configs.

Additional information

Verified no regression on existing single-op perf tests.
The ADD operator shows same runs-per-loop and memory-per-run on both branches, with negligible performance diff.

master (389ff61):

ADD(type=f32,ne=[4096,1,1,1],nr=[1,1,1,1],nf=1,perm1=0,src_overlap=0):              343980 runs -     2.95 us/run -       48 kB/run -   15.50 GB/s
ADD(type=f32,ne=[4096,1,1,1],nr=[1,512,1,1],nf=1,perm1=0,src_overlap=0):              5464 runs -   194.47 us/run -    24576 kB/run -  120.52 GB/s

this PR:

ADD(type=f32,ne=[4096,1,1,1],nr=[1,1,1,1],nf=1,perm1=0,src_overlap=0):              335790 runs -     2.98 us/run -       48 kB/run -   15.37 GB/s
ADD(type=f32,ne=[4096,1,1,1],nr=[1,512,1,1],nf=1,perm1=0,src_overlap=0):              5464 runs -   195.81 us/run -    24576 kB/run -  119.70 GB/s

Requirements

• I have read and agree with the contributing guidelines
• AI usage disclosure: NO