pass recursive lambda self param by reference

[cameroncuster]

Changes the self parameter of every self-passing recursive lambda from auto (by value) to auto&& (by reference) — 20 lambda sites across 18 files.

Why

auto dfs = [&](auto dfs, int u) { ... };   // before: closure copied into every frame
auto dfs = [&](auto&& dfs, int u) { ... }; // after: one closure, 8-byte ref per frame

A [&] closure is one word per captured entity. With the by-value form, every recursive call copies the whole closure into the new stack frame (and it must stay live across the recursive call, so it can't be elided). With auto&&, each frame holds only a reference. Call sites are unchanged.

Measured impact

Benchmark: recursive lambda capturing 4 locals (same shape as this library's DFS lambdas), measuring the address delta of a local between consecutive recursion depths. GCC 16, arm64 macOS:

	auto dfs (self by value)	auto&& dfs (self by ref)
frame size at -O2	136 B	40 B
frame size at -O0	112 B	64 B
stack at depth 10^7	~1.3 GB	~380 MB

The closure copy is 32 B, but it forces additional spills/alignment on top, tripling the frame at -O2. Exact numbers are ABI/capture-count dependent (x86-64 judges will differ), but the direction and rough ratio are robust.

Practical consequences:

• On a default 8 MB stack (running locally without ulimit -s), this is the difference between crashing at depth ~60k vs ~200k on a path graph.
• On judges where stack = memory limit, a 10^6-node path graph costs ~130 MB of stack by value vs ~40 MB by ref.
• Runtime speed is unchanged — the copies are trivial and inlined; this is purely a stack-footprint improvement.

Cost

+2 characters per lambda. No behavior change; all call sites (dfs(dfs, ...)) are identical.

If the library later moves to C++23 deducing this, the equivalent recommended form is this auto&& dfs.

Verification

• clang-format --dry-run --Werror (dev.clang-format) passes on all 18 files
• All affected Library Checker/Aizu tests compile clean with g++ -std=c++20 -O2 -Wall -Wextra (lca, edge_cd, centroid decomp, kth_path x4, shallowest, subtree isomorphism, scc x2, offline incremental scc, euler walk, two-edge cc, bcc x2, min-plus convolution)