tests: benchdnn: print csv friendly perf header

[jondea]

Description

Most of the benchdnn perf-templates already emit CSV-shaped data, but there are no headers, and the template doesn't match the columns. Specifically, %DESC% can expand to multiple columns for some drivers, which makes the output hard to parse. This change adds a CSV header row that matches each driver's concrete %DESC% expansion while preserving the existing Output template: ... banner. This makes the CSV output easy to consume with standard tools (e.g. xsv).

E.g. after this change,

./tests/benchdnn/benchdnn --conv --mode=P --perf-template=csv  mb1_ic1oc1_id1ih5iw5_od1oh3ow3_kd1kh3kw3_sd1sh1sw1_pd0ph0pw0_dd0dh0dw0 \
    | grep '^perf,' \
    | xsv table

perf  engine  impl         name  dir    sdt          stag  wtag  dtag  alg     attr  g   mb  ic  id  ih  iw  oc  od  oh  ow  kd  kh  kw  sd  sh  sw  pd  ph  pw  dd  dh  dw  Gops      max_ctime  min_time  min_Gflops  avg_time  avg_Gflops
perf  cpu     jit:sve_128        FWD_B  f32:f32:f32  any   any   any   direct        1   1   1   1   5   5   1   1   3   3   1   3   3   1   1   1   0   0   0   0   0   0   1.62e-07  0.994385   0.137695  0.00117651  0.32537   0.000497895

Not sure how to test this using ctest, I'm open to ideas...

Checklist

General

• Do all unit and benchdnn tests (make test and make test_benchdnn_*) pass locally for each commit?
• Have you formatted the code using clang-format?

[vpirogov]

I wonder if we actually need information included into problem descriptor (%DESC%) split into multiple columns. If the only goal is getting 1-1 correspondence between column headers and data I would argue for printing %DESC% in the same format as used in benchdnn input. Something like this in your example:

Output template: perf,%engine%,%impl%,%name%,%dir%,%sdt%,%stag%,%wtag%,%dtag%,%alg%,%attr%,%DESC%,%Gops%,%+ctime%,%-time%,%-Gflops%,%0time%,%0Gflops%
perf,cpu,jit:avx512_core,,FWD_B,f32:f32:f32,any,any,any,direct,,mb1ic1id1ih5iw5oc1od1oh3ow3kd1kh3kw3pd0ph0pw0,1.62e-07,0.689209,0.00146484,0.110592,0.00220769,0.07338

This approach has a benefit of keeping the same number of columns between different primitives in CSV format.