iterative Tarjan SCC to avoid stack overflow on large graphs

[LdDl]

recursive Tarjan risks uncontrolled Go stack growth on city-scale graphs with deep DFS chains (100k+ vertices). Iterative version bounds memory explicitly so I've replaced recursive strongConnect with iterative version using explicit call stack:

• Pre-build flat adjacency list (CSR-style) - single allocation instead of per-vertex map iteration
• Reuse call stack buffer across strongConnect invocations

Metric	Before	After	Delta
ns/op	36,351,165	36,121,026	-0.6%
B/op	15,507,960	19,277,907	+24%
allocs/op	8,058	8,198	+1.7%

graph has 40k vertices: time and allocs are changed barely. Also +3.8MB is the one-time cost of the flat adjacency buffer - negligible compared to the graph itself.