Adds generic GPU TrackEvent support across Perfetto

[dreveman]

Summary

Adds generic GPU timelines built from standard TrackDescriptor and
TrackEvent packets. Producers can control GPU hierarchy, naming, ordering,
annotations, counters, and state tracks without adding API-specific
hierarchy fields.

GPU hierarchy uses ordinary parent_uuid relationships. Process-scoped
concepts such as CUDA streams are associated with a process by parenting the
GPU root to a process TrackDescriptor.

Render-stage integration

GpuTrackDescriptor can bind an exact authored track to either:

oneof render_stage_queue {
  uint64 hw_queue_iid = 2;
  uint64 logical_queue_id = 3;
}

Matching GpuRenderStageEvents are projected onto that authored track while
remaining available as canonical gpu_slice rows. Authored annotations and
projected render-stage events use normal TrackEvent sibling merging:

• tracks merge by name by default;
• sibling_merge_key provides an explicit merge identity;
• SIBLING_MERGE_BEHAVIOR_NONE keeps the sources separate.

This lets producers add annotations directly to hardware or logical queue
tracks. For example, a producer can emit iteration ranges, scheduling
phases, queue-saturation intervals, or instant markers on the bound
descriptor. Annotations are displayed on the same visual track as the
matching GPU work, without duplicating render-stage events in the producer.

Flow relationships are mirrored onto projected representations without
creating cross-hierarchy or Cartesian duplicate flows.

Hardware queue IDs are scoped to the trusted packet sequence. Logical queue
IDs are scoped to the process resolved from the render event's graphics
context.

UI behavior

A global GPU descriptor introducing a concrete gpu_id acts as the
placement anchor for that GPU. Existing frequency, memory, counters, and
other GPU-specific tracks become direct children of the anchor. This avoids
synthetic structures such as Counters -> GPU N.

For example, a producer can create:

NVIDIA A100
  Frequency
  Memory
  Utilization
  Channels
    Channel #1
    Channel #2

Process-owned hierarchies can similarly represent logical concepts:

CUDA
  Device #0
    Stream #2

Only successfully bound hardware or logical queues move into authored
locations. Unbound sources remain in the existing fallback hierarchy, so
authored and inferred tracks can coexist.

Traces without generic GPU descriptors retain the current UI hierarchy,
naming, and ordering.

Other changes

• Exposes generic GPU scope and process ownership in Trace Processor.
• Keeps GPU placement policy in the GPU plugins while TrackEventPlugin
  remains generic.
• Resolves projected GPU Compute selections back to canonical slices.
• Updates visualization SQL, generated Python/Rust bindings, tests, and GPU
  documentation.
• Bumps perfetto-protos-gpu to 1.0.3.

[github-actions]

🎨 Perfetto UI Builds

• c40a559 (2026-07-30 14:20 UTC): https://storage.googleapis.com/perfetto-ci-artifacts/gh-30550408133-1-ui/ui/index.html
• c40a559 (2026-07-30 16:11 UTC): https://storage.googleapis.com/perfetto-ci-artifacts/gh-30550408133-2-ui/ui/index.html
• 16a87d6 (2026-07-31 17:53 UTC): https://storage.googleapis.com/perfetto-ci-artifacts/gh-30652655425-1-ui/ui/index.html
• d6b440d (2026-07-31 18:20 UTC): https://storage.googleapis.com/perfetto-ci-artifacts/gh-30654253741-1-ui/ui/index.html
• 15de89d (2026-08-04 17:45 UTC): https://storage.googleapis.com/perfetto-ci-artifacts/gh-30934991841-1-ui/ui/index.html

[dreveman]

@LalitMaganti You might want to look at the individual commits as the full set of commits does a lot. This PR provides a complete picture but we can land each commit as a separate PR if you prefer.

[LalitMaganti]

Still out on vacation for another week and a bit so will take a look when I'm back.

[dreveman]

Still out on vacation for another week and a bit so will take a look when I'm back.

No worries. I iterated on this approach some and I think the current version is now something I'm pretty happy with and provide all the features I was missing from perfetto when it comes to GPU workloads.