test(b+tree): make the absl::btree_set comparison buildable, and widen it

[psiha]

The benchmark already had an absl::btree_set arm behind #define HAVE_ABSL 0, with no way to turn it on. -DPSI_VM_BENCH_ABSL=ON now fetches abseil through CPM and defines it. absl::btree_set is the closest shipping container to this one — it also redistributes into a sibling before splitting — so it is the right thing to be measured against.

It also measures more than it did: sorted bulk insertion is every container's best case and is not the state a long-lived index is in, so there is now a one-by-one random insertion arm and a footprint column for both containers (nodes_reserved() here; a counting allocator for absl, which exposes bytes_used() only on the internal btree).

What it says

7.65M int, one box, Release + LTO. abseil's default node is 256 bytes (/*TargetNodeSize=*/256), so the 4096 column is a 16× geometry mismatch and the 512 column a 2× one.

	psi::vm 512B	absl 256B		psi::vm 4096B	absl 256B
sorted bulk insert	4 ns	7 ns	1.75× faster	4 ns	7 ns	1.75× faster
lookup after bulk	208	176	1.18× slower	284	188	1.51× slower
random one-by-one insert	163	161	parity	329	181	1.82× slower
lookup after random	202	206	parity	354	217	1.63× slower
bytes/key, bulk	4.30	4.33		4.04	4.33
bytes/key, random	4.91	5.07		4.49	5.07

At the default 512-byte geometry this container holds its own: parity on random insert and random lookup, 1.75× faster on sorted bulk insert, slightly smaller. The one loss is lookup on a freshly bulk-built — fully packed — tree.

At PSI_VM_BT_PAGE_SIZED_NODES it does not: 1.5–1.8× slower on lookups and one-by-one inserts, and 2× slower than itself at 512 bytes on random insertion. The likely mechanism is cache lines touched per node: 256 B is 4 lines and 512 B is 8, against 64 for a 4096-byte node — fewer tree levels, far more lines per descent.

⚠ Caveat worth stating rather than burying: with int/std::less, 512 bytes falls under linear_search_byte_limit and 4096 does not, so the two columns differ in node size and in intra-node search. Each is the right choice for its own geometry, so this is a fair end-to-end configuration comparison — but it is not an isolated node-size measurement.

Build note

psi.build's MSVC toolchain appends __GNUC__ to the global compile definitions (clang-cl does support the GNU-style extensions), and abseil reads that as "GNU platform" and builds its libsupc++ demangler, which wants cxxabi.h. The definition is dropped for abseil's subtree only, by filtering the directory property before CPMAddPackage — a subdirectory inherits it as it stands at the add_subdirectory() call.

Stacked on #159.