[native] Deduplicate BindGroup usage tracking per pass

In a render pass, render bundle or compute pass, rebinding the same
BindGroup across draws or dispatches (e.g. with varying dynamic offsets)
repeatedly walks the bind group layout and records identical
whole-resource usage on every SetBindGroup call.

Add a 64-bit PassTrackerID TypedInteger generated atomically by
DeviceBase (GetNextPassTrackerID) and give one to each pass-level
tracker (RenderPassResourceUsageTracker and
ComputePassResourceUsageTracker). Track the last pass tracker a
BindGroup was recorded in intrusively on BindGroupBase via an
atomic<PassTrackerID> checked in ShouldAddToPassTracker() with relaxed
atomics. Duplicate additions caused by concurrent use in multiple
encoders are benign.

- Render passes and bundles:
  RenderPassResourceUsageTracker::AddBindGroup skips bind groups
  already recorded in the pass.
- Compute passes:
  ComputePassResourceUsageTracker::AddResourcesReferencedByBindGroup
  skips bind groups already referenced in the pass. Per-dispatch
  synchronization scopes still record every bind group since each
  dispatch needs its own SyncScopeResourceUsage.

To keep a single, type-safe code path, the non-deduplicated walk
becomes the protected SyncScopeUsageTracker::AddBindGroupUsages(), and
a new DispatchResourceUsageTracker exposes it as AddBindGroup() for the
per-dispatch scopes of compute passes.

On Animometer (10,000 draws/frame), this reduces AddBindGroup CPU cycle
cost from 7.00% to 0.71% (~220 us saved per frame) on Adreno 840.
Gained 8.8% CPU FPS. With 1,000 to 10,000 dispatches per frame that
rebind the same bind group with dynamic offsets, this reduces Dawn CPU
instructions per frame by 4-6% on Pixel 11.

Bug: b/564314926, b/570398140
Change-Id: I682bacceafa6d1de156e0b61669d8f814303d17a
Reviewed-on: https://dawn-review.googlesource.com/c/dawn/+/342775
Reviewed-by: Brandon Jones <bajones@chromium.org>
Reviewed-by: Corentin Wallez <cwallez@chromium.org>
Commit-Queue: Sanjeevi Suresh <sanjeevisuresh@google.com>