ui: Fix GpuByProcess plugin and nest tracks under process groups (#5789)

The previous GpuByProcess plugin assumed each process had a single hw
queue on a single GPU and rendered all GPU slices on one combined track
per process. When a process spanned multiple hw queues or multiple GPUs,
slices from different queues collided at depth 0 on that combined track
and got dropped or mis-rendered by the slice mipmap, leaving render
stage events visible on the global hw queue tracks but missing from the
per-process view.

Fix this by partitioning per-process events by (upid, hw_queue_id) and
registering one SliceTrack per tuple. hw_queue_id is per-sequence so two
processes can reuse the same value for distinct queues; the (upid,
hw_queue_id) tuple disambiguates them. Sub-track names mirror the global
hw queue track names ("Channel #1", etc.).

Nest the resulting hw queue tracks under a per-process "GPU" summary
group, attached to the process group created by ProcessThreadGroups (or
created lazily if no group exists yet). When a process has events on
more than one GPU, insert an additional per-GPU layer ("GPU 0", "GPU 1",
...) mirroring the layout used by dev.perfetto.Gpu under the global GPU
section. Tracks and groups are only created for (process, queue)
combinations that actually have events.