ui: render sampled stacks as a per-thread flamechart track

Sampled call stacks are currently only visible as one instant per
sample or as an aggregated flamegraph on selection; there is no way to
see how stacks evolve over time. Replace the per-thread callstack
samples track with a flamechart track: row 0 keeps one diamond marker
per sample (click to see its callstack, like the instants view it
replaces) and the rows below render the maximal-run decomposition
computed by _flamechart_runs! over the inline-expanded callstack
forest. All stack-sample markers (including the process callstacks
tracks) render as diamonds.

The flamechart only applies when the session's timebase approximates
wall time (cycles, cpu-clock, ...); event-count timebases keep the
instants view. Threads with several sampling sessions get one track
per session under a summary track, preserving per-session area
selection. All expensive work happens on the first render of the
track, so trace load pays no flamechart cost.

Frames are colored by origin, consistently across all stack-sample
surfaces: the process's own binary, shared libraries and the kernel
each get a distinct muted hue and unresolved frames are grey like the
flamegraph. Bars carry a dashed bottom hairline (RECT_PATTERN_SAMPLED),
implemented in both the WebGL shader and the Canvas2D fallback, and
tooltips mark durations as "(sampled)". Empty frame names display as
"unknown".

Stack depth is one dial ('Stack frames': No frames / 4 frames / All
frames): hidden frames sit behind a clickable grey "…" row which
steps the view up one level, so the flamechart stays discoverable from
the markers-only state. Threads with instrumented slice tracks default
to markers only; other threads default to the 4-frame peek, and only
the busiest few reveal themselves so system-wide traces with hundreds
of sampled threads stay readable.

Track naming drops the source prefix when only one source emits stack
samples, per-session splits only happen with at least two sessions,
and process callstacks tracks are skipped when they would duplicate a
single sampled thread (the common shape for kernel threads).