ck-libs/cache: CkTreeCacheManager, a process-shared tree-node cache with CkCache's protocol

[lvkale]

Adds CkTreeCacheManager<Key, Traits> to ck-libs/cache, beside CkCacheManager. It keeps CkCache's request/reply protocol (requestData, recvData, chunk finish) but stores fetched tree nodes once per process instead of once per PE: one process-shared store owned by the rank-0 branch, lock-free on the fetch path (atomic child-pointer exchange and parked request lists, from paratreet2's TreeCacheCore.h, vendored here with a provenance block). The application supplies a small static Traits class that maps its node type onto the store. First client: ChaNGa's tree-node cache (N-BodyShop/changa#228), where a configure option selects this or CkCacheManager.

Changes to existing code: CkCache.h now always compiles its request/reply counters (a few integer increments per request). Before, they were compiled only under the application macro COSMO_STATS, which changed the class layout in the application but not in libmoduleCkCache (the registration stubs live in the library). CmiResetMaxMemory is skipped under CMK_RECONVERSE. Nothing else in charm uses the new class.

Validation:

• Classic Charm++ (netlrts-darwin-arm8-smp, and mpi-linux-x86_64-smp on Purdue Anvil) and reconverse (Anvil, NCSA Delta, OLCF Frontier): ChaNGa teststep energy and restart tests pass, results identical to CkCache.
• ChaNGa gravity, 80.6M particles, big steps 1-3 summed, new cache vs CkCache in the same tree (--enable-smpcache=no), 2 repeats:
  • classic mpi-smp on Anvil, 8 processes x 8 PEs per node: 0.79 at 4 nodes, 0.77 at 8 nodes (cache teardown per step 2.5-3.6 s -> 0.2-0.55 s).
  • reconverse: 0.82-0.95 across 17 layouts on Delta; 0.83-0.96 at 4-32 nodes on Frontier with the simulation's own parameters.

The same commit applies cleanly to reviewed-with-reconverse (branch treecache-nodecache, needs a rebase onto its current tip).

🤖 Generated with Claude Code

[lvkale]

The same commit for reviewed-with-reconverse is #4014.