Bump reconverse to 988a8b4 (#267, #269 randomized msgq, #272 PE queue counter)

[lvkale]

Bumps contrib/reconverse from 2e5775d83 to 988a8b4 (reconverse main), bringing in:

• reconverse Destructors for NodeGroup classes will deregister from the Group table #267: self-sends go to the PE's priority queue; the old scheduler bounds its thread-queue streak.
• reconverse Arguments to an sdag method are not available for defining condition of a while statement #269: randomized message queue, +randomized_msgq [+randomized_seed N] at runtime (prints the seed; priorities, FIFO order and expedited delivery are not respected by design).
• reconverse Memory leaks in Charisma RTS code #272: atomic length counter on the PE and node queues; opt-in RECONVERSE_USE_JEMALLOC; tests/pingpong_after_fanin.

Charm's sources use none of the changed internal queue API (no CmiTakeSelfQueue, pollConverseThreadQueue or ConverseQueue references; reconverse's include/ is unchanged between the two commits), so no charm change is needed. #4029 (removing charm's build-time randomized option, superseded by the runtime flag) can merge after this.

Checked on macOS (reconverse-darwin-arm8 --with-production): build clean; benchmarks/charm++/pingpong +pe 2 all 19 phases; tests/charm++/megatest +pe 4 all 54 tests. With +randomized_msgq, megatest aborts in priotest ("message received out of order"), which is the expected consequence of randomized delivery (see charm #4032 for the delegation/multicast case).

🤖 Generated with Claude Code

[lvkale]

Full single-process sweep on macOS of every directory under tests/, examples/ and benchmarks/ with a test target (254 leaf directories: 182 run, 72 skipped for AMPI/CUDA/f90/MPI-interop/ParFUM/POSE/etc.), this branch against the same charm commit with the old submodule (2e5775d83): identical results in every directory. 154 pass, 9 fail, 1 timeout, 18 compile failures in both trees, all pre-existing (communication_overhead segfault at 64 KB, xcastredn CmiMaxReductions abort, hello/local segfault, jacobi2d-iter shrink-expand timeout, io_read segfault, randomTRAM3D delivery abort, three nodegroup zero-copy tests and partitions that need more than one process; the compile failures are stale test code against removed interfaces, missing headers, and libc++ rejections). No regression from the bump.