tests: add renderer performance infrastructure

[tritao]

Builds on #55.

Purpose

The retained and LegacyGL renderers need to be compared using the same scenes
and the same definitions of work. Correctness tests alone cannot show whether
time moved from traversal into planning, submission, GPU completion, picking,
or selection, while wall-clock thresholds on ordinary CI runners would be too
noisy to trust.

This PR adds shared, deterministic workload and measurement infrastructure.
It is the foundation used by the performance layers above it; it does not
optimize renderer behavior itself.

What changed

• Add deterministic CPU and hardware OpenGL benchmark executables with JSON
  output.
• Compare LegacyGL, DrawList compatibility, and DrawList core paths where the
  workload is semantically supported by each.
• Measure retained traversal/construction, planning, backend preparation,
  submission, GPU completion, picking, and selection phases separately.
• Share workload construction between benchmarks, parity tests, and an
  interactive viewer.
• Add reusable visible/resizable GL test sessions and portable framebuffer
  operations through Coin's GL glue.
• Register capability-based benchmark and stress CTest labels.

How results should be used

Timing is informational and intended for controlled hardware comparisons.
Ordinary correctness CI compiles the tools and runs smoke coverage, but does
not fail a PR for small wall-clock changes. Deterministic work counters,
checksums, and parity assertions protect against empty or semantically
different fast paths.

Validation

• CPU benchmark, hardware GL benchmark, parity, and viewer smoke coverage.
• Compatibility and core DrawList contexts, plus LegacyGL where enabled.
• Portable shared-library test boundaries on Windows.
• Local compatibility build with 31 selected unit, GL, viewer, parity, and
  benchmark tests; full platform coverage is exercised by the top stack PR.