Fix benchmarks

[jordanrule]

Fix benchmark workflow trigger model + collector reliability. Fixes: #991

This PR fixes two benchmark CI issues:

1. Benchmarks no longer run on a 15-minute schedule

   • Removed schedule trigger from .github/workflows/benchmarks.yml
   • Benchmarks now run per-commit (push to trunk) and via workflow_dispatch
   • Keeps concurrency.cancel-in-progress: true so newer commits supersede stale runs

2. Benchmark collection path is no longer brittle

   • Checkout now pins to ${{ github.sha }} (instead of floating trunk) in benchmark and collector jobs
   • Added explicit failure when expected benchmark artifacts are missing
   • Fixed platform extraction from artifact names:
     • benchmark-results-<sha>-x86-64-linux now maps to x86-64-linux (previous logic collapsed to linux)
   • Fixed shell syntax in "Commit and push to perf branch" (missing fi)
   • Removed schedule-specific decision branches/messages that no longer apply

Local verification performed

• YAML validation:
  • ruby -e 'require "yaml"; YAML.load_file(".github/workflows/benchmarks.yml"); puts "YAML parse OK"'
  • Result: YAML parse OK
• Platform parsing sanity check:
  • Verified expected extraction for x86-64-linux, aarch64-linux, aarch64-macos
• Attempted end-to-end benchmark run:
  • ./bench.sh --iterations 1 --output /tmp/rue-bench-results.json --no-history
  • In this local environment, the run stalled in buck2 build before producing results JSON

Why this should fix the CI failures

• The collector now matches artifact names deterministically per commit SHA.
• Platform keys now match website/perf history file naming conventions.
• Broken shell block no longer causes collector step failure.
• Trigger behavior is aligned with per-commit benchmarking (no redundant schedule path).