Test suite cleanup: separate benchmarks from tests, fix silently-passing cases

[Nikos-d]

Scanned all 19 registered test targets (plus the disabled test_gjk_epa.cpp) for tests that silently pass, benchmarks living inside TEST-registered CTest binaries, and other test-hygiene issues. Found during work on #nikos/tight_validation and split out for a dedicated cleanup PR.

1. Silently-passing test cases (zero assertions)

Catch2 reports a TEST_CASE with no CHECK/REQUIRE as passed by default (verified against this project's Catch2 v2.13.10 build) — these currently give no signal at all:

• world/test_broadphase.cpp:1208 — "Optimization benchmark": only std::couts results, no CHECK
• world/test_broadphase.cpp:1364 — "Test": same, prints stats only. Also a terrible name regardless.
• world/test_primitives.cpp:121 — "Collision method benchmarks": pure BENCHMARK(), no CHECK
• world/test_gjk_epa.cpp:190 — "Benchmark - GJK computation time vs point cloud size": same (file isn't even built today, see Andre/ur5 #3)

2. Benchmarks living inside TEST-registered CTest binaries

Should move to tests/benchmarks/ as proper nanobench BENCH targets (see tests/benchmarks/bench_core.cpp for the existing pattern), so they never run under ctest or a default test invocation:

• world/test_broadphase.cpp — 3 of its 4 TEST_CASEs are benchmarks: the two above, plus "Benchmark - Broadphase vs Blast..." at line 626. That one's a hybrid: 3 BENCHMARK() blocks + one real CHECK(is_close(...)) at the end comparing naive vs. BVH distance — the check needs to stay behind as a TEST when the benchmarks move out.
• world/test_primitives.cpp:121 — "Collision method benchmarks", pure BENCHMARK().
• world/test_broadphase.cpp:1119 — "Constraints calculation" isn't a benchmark by name, but it's gated by a hardcoded #define BENCHMARKING_ON 0 (line 7) that makes it loop num_tests_per_world = 1e3 per world instead of using BENCHMARK(). This is why it hangs past 150s when run directly. It has real CHECK(is_close(...)) assertions cross-checking with_segments/broadphase/double_broadphase, so it should stay a TEST, but the iteration count needs fixing independent of the benchmark move — 1000 iterations × 8 worlds × 3 solve methods is not a reasonable default ctest cost.

3. Dead/disabled test infrastructure

• world/test_gjk_epa.cpp — entire file commented out of tests/CMakeLists.txt (line 105), so it doesn't build or run at all today. Has 7 commented-out BENCHMARK() blocks inside too (lines 159-244). Needs a decision: revive it cleanly (split test/bench like everything else) or delete it — right now it's dead weight nobody notices.
• world/test_primitives.cpp:144,156 — a commented-out BENCHMARK() and an entire commented-out TEST_CASE("Collision method comparison exhaustive (Box-capsules)").
• Dead #define CATCH_CONFIG_ENABLE_BENCHMARKING (enables the macro but the file never calls BENCHMARK()) in: optimization/test_task.cpp, trajectory/test_bsplines.cpp, optimization/test_optimization_task_violation.cpp, world/test_gjk_epa.cpp. Harmless but noise — cheap one-line deletes.

4. Non-determinism / flakiness risk

• world/test_broadphase.cpp:629-632 — "Benchmark - Broadphase vs Blast..." picks its world via unseeded std::random_device/std::mt19937 every run. Moot if it moves to BENCH, but if the trailing CHECK(is_close(...)) stays behind as a TEST, it should get a fixed seed so failures are reproducible.
• world/test_gjk_epa.cpp — "Test - GJK Box vs Capsule (random generation)" is randomized by its own name; same concern, moot while the file's disabled.

5. CTest/Catch2 config gap

• tests/CMakeLists.txt:63 — add_test(NAME ${target}_test COMMAND ${target}) runs each binary with no arguments, so -w NoAssertions is never passed. That's why item Andre/benchmark #1 can happen silently — worth adding at the CMake level so this class of bug becomes self-reporting instead of needing another manual scan like this one.