fix(ci): render all benchmark columns in summary, not just allowlisted backends

[hannahli-nv]

Description

The benchmark summary page silently dropped columns for any backend not in the hardcoded allowlist {CuTile, PyTorch, Triton, TorchCompile} in .github/scripts/format_benchmark_summary.py. Non-allowlisted keys were also overwriting param_name, so the x-axis column was lost in several benches.

This swap replaces the allowlist with direct iteration over configs[0].keys(), which run_all_json.py already emits in Triton's perf_report order (x-axis params first, then backends).

Before / After (from CI run 24546853000 artifacts)

bench_attention_backward (4 SDPA backends + N_CTX)

• Before: | SDPA-Math | CuTile | (N_CTX / SDPA-Flash / SDPA-MemEff dropped)
• After: | N_CTX | CuTile | SDPA-Flash | SDPA-MemEff | SDPA-Math |

bench_bmm (multi-axis M/N/K)

• Before: | K | CuTile | PyTorch | (M / N dropped)
• After: | M | N | K | CuTile | PyTorch |

bench_mix_triton_cutile

• Before: | Triton+CuTile | PyTorch | (N dropped)
• After: | N | PyTorch | Triton+CuTile |

Every bench already rendering correctly (e.g. the ~22 single-x CuTile + PyTorch files): unchanged.

Diff

14 insertions, 17 deletions in a single file. No schema changes, no new imports, no changes to run_all_json.py or check_benchmark_regression.py.

Test plan

• Patch rendered against real CI artifacts for bench_attention_backward, bench_bmm, bench_mhc, bench_mix_triton_cutile, and bench_swa_attention — all produce expected tables.
• Every bench whose keys are {x, CuTile, PyTorch} continues to render identically to today.
• Next tilegym-ci run's summary page shows all backend columns for the affected benches.

Known limitations (deliberately not addressed here)

• run_all_json.py::parse_benchmark_output still tokenizes the pandas header with .split(), so any future bench whose line_names entry contains whitespace would fragment into multiple columns. Hardening this parser is left to a follow-up.
• check_benchmark_regression.py has a similar allowlist, but uses break, so it already picks the correct param_name in practice. Leaving it alone to keep this diff focused.

Effect

CI Configuration

config:
  build: true
  # valid options are "ops", "benchmark", and "sanity"
  test: ["benchmark"]

Checklist

• Code formatted and imports sorted via repo specifications (./format.sh)
• Documentation updated (if needed)
• CI configuration reviewed

[copy-pr-bot]

This pull request requires additional validation before any workflows can run on NVIDIA's runners.

Pull request vetters can view their responsibilities here.

Contributors can view more details about this message here.

[hannahli-nv]

/ok to test 374a9f4

[hannahli-nv]

/ok to test 60a2145