The GPU time reported by the WebGPU profiler (and shown in mini-stats) was massively overestimated, growing with the number of render/compute passes even when the GPU was mostly idle. On tile-based / pipelined GPUs (Apple Silicon is a well known example) consecutive passes overlap in time — the vertex stage of one pass runs concurrently with the fragment stage of the previous one — so summing per-pass timestamp durations double-counts the overlap.

This reports the frame GPU time as the span between the earliest begin and latest end timestamp across all passes, which stays within the real frame time. Per-pass timings are kept as-is (useful for relative comparison) but are now clamped to be non-negative, fixing the negative readings some users observed.

Changes:

• WebGPU frame GPU time is now the timestamp span (max end − min begin) across all passes, instead of the sum of per-pass durations.
• Per-pass GPU durations are clamped to ≥ 0, fixing occasional negative timings caused by begin/end timestamps reported out of order by the driver.
• WebGL profiling is unchanged — it already measures the whole frame with a single query, so its reported time does not overlap.
• Fixed a debug-label bug where the timestamp staging buffer was mislabelled.

Closes #8350