Addresses #523

Summary

The current flamegraph output stacks all functions linearly by descending
inclusive cost — a design described as a quick
cost overview using inferno as a rendering backend.

This works as a sorted bar chart, but the flamegraph y-axis (call depth)
carries no meaning — sibling calls appear nested rather than branching. Since
callgrind output already contains caller-callee edges (cfn=, calls=,
per-call-site costs), we can do better.

This PR restores caller-callee edge tracking (present in 190dfd6, removed in
1f2e0bd) and uses DFS from the sentinel/root to emit folded stacks that
reflect actual call-graph structure.

Before / After

Given parent of total cost 90 Ir from calling inline_me (10 Ir), work_a (60 Ir), work_b (20 Ir):

fn parent() {
    inline_me(); // 10 Ir
    work_a(); // 60 Ir
    work_b(); // 20 Ir
}

Before (linear staircase):

above_parent 12345
above_parent;parent 90
above_parent;parent;work_a 60
above_parent;parent;work_a;work_b 20

After (branching call tree):

parent; 10
parent;work_a 60
parent;work_b 20

Now, parent is annotated with its own cost in the output graph. Rendering the graph compiles its total cost, so that remains available when viewing the graph.

Full example: https://github.com/turbocrime/gungraun-flamegraph-example

Github disables the fancier features of the SVG (interactivity, etc) when embedded, so please run the repro to get the full idea.

What changed

• hashmap_parser.rs: parse_with_edges() exposes caller-callee edges via
  callback during parsing
• flamegraph_parser.rs: replaces BinaryHeap + .windows(2) with a
  CallGraph struct and DFS stack emission; sentinel is used as DFS root when
  present
• New test fixtures for branching and recursive call graphs

Limitations

Inferno forces lexical sorting when rendering the graph.