library: C++23 — deducing this for some recursive lambdas

[cameroncuster]

Converts the library to C++23. Companion to #228 (infra/tests std bump), which this PR is stacked on — it must merge after #228, since the code here needs -std=c++23 to compile.

Deducing this for recursive lambdas (15 lambdas, 12 files)

Recursive lambdas in library/ converted from the pass-self-explicitly idiom:

// before
auto dfs = [&](auto&& dfs, int u) -> void { ... dfs(dfs, v) ... };
dfs(dfs, 0);
// after
auto dfs = [&](this auto&& dfs, int u) -> void { ... dfs(v) ... };
dfs(0);

Justification (correct/efficient/readable):

• Readable: removes the self-parameter and the doubled first argument at every recursion and call site — the most confusing idiom in the codebase for newcomers, and strictly shorter.
• Correct: no way to accidentally pass the wrong callable as the self argument.
• Efficient: no change; same calls after inlining.

Converted files: trees/{centroid_decomp ×2, edge_cd ×2, shallowest_decomp_tree, uncommon/subtree_isomorphism}, graphs/{bcc_callback, euler_path, scc, uncommon/{bridges, cuts}, strongly_connected_components/{offline_incremental_scc, add_edges_strongly_connected}}, convolution/min_plus_convolution_convex_and_arbitrary.

Not converted (GCC bug): the 8 lambdas inside struct constructors — trees/{hld ×2, tree_lift, lca_rmq, linear_lca, uncommon/{hagerup_kth_par, ladder_decomposition, linear_kth_par}} — access data members through the captured this. GCC 14/15 rejects member access inside a deducing-this lambda that captures the enclosing this (GCC PR113563, fixed in GCC 16). CI's g++-14 (same as the Kattis judge) fails on them, so they keep the pass-self idiom until the toolchain moves past the bug.

library/contest/tester.sh: -std=c++20 → -std=c++23

The stress tester should compile with the same standard as the contest judge (SEUSA 2025 on Kattis: g++-14 -std=gnu++23).

Evaluated and rejected: std::generator for get_prime_factors.hpp

Measured before deciding. Rewriting the factor loop as a std::generator<int> coroutine:

• 2.7× slower — factorizing all of [2, 1e7): loop idiom 160 ms vs generator 430 ms (g++ -O2), plus a heap allocation per call (coroutine frame)
• Trips -Werror — GCC's own <generator> triggers -Wmismatched-new-delete under -O2 -Wall (known false positive, GCC PR109224, still open), which would fail compile_gcc CI and require a suppression for zero gain

The existing "loop header you complete with a body" idiom stays.

Validation

• CI compile_gcc (g++-14) confirmed the constructor-lambda failures; those 7 files reverted, everything else passes
• All test files covering the changed headers compile clean with g++ -std=c++23 -O2 -Wall -Wextra
• clang-format --dry-run --Werror passes on all changed files
• Grep sweep: zero remaining NAME(NAME, ...) self-passing call sites in the converted files
• Behavior is identical (pure syntax transformation); the full verifier run in CI on top of tests/infra: move to C++23 (std pins, g++-14 in CI, deducing this in test code) #228's toolchain is the final gate