fix(hip): release Q8 memo before pool teardown

[dusterbloom]

Summary

Release LUCE_Q8_MEMO allocations before the CUDA/HIP context destroys the pools that own them.

This keeps backend shutdown on the normal resource-lifecycle path instead of aborting in the legacy HIP pool destructor. It changes no inference kernels, model backends, cache formats, or generated values.

Root cause

ggml_backend_cuda_context::luce_q8_memo stores ggml_cuda_pool_alloc owners. The context declares its pool array after the memo vector, so C++ member destruction destroys the pools first. A memo allocation retained after the final graph evaluation therefore remains checked out when the legacy pool destructor checks pool_size == 0.

On gfx1151, that shutdown assertion aborted the HIP process. The abnormal exit could poison the KFD workqueue and prevent the next HIP process from initializing until reboot.

Changes

• Drain Q8 memo entries in LIFO order from ggml_backend_cuda_context's destructor while its pools are still alive.
• Preserve the LIFO release requirement of VMM pools.
• Add a CUDA/HIP regression test that performs a Q4_0 matvec with LUCE_Q8_MEMO=1 and then frees the backend with the memo allocation still retained.

Relationship to the stacked DS4 PRs

This fix was checked against the complete open stack #502 -> #503 -> #496:

• feat(ggml): vendor ROCmFPX quant support for DeepSeek V4 Flash #502 changes ggml-cuda.cu for ROCmFPX types, DS4 HC/SwiGLU, and dispatch, but does not change the context destructor or allocation ownership.
• fix/feat(ds4): ROCmFPX byte-safe decode + fused-decode path + ds4 AR optimizations #503 and feat(ds4): add DSpark speculative runtime + optimizations  #496 do not change the CUDA/HIP context destructor.
• At the head of feat(ds4): add DSpark speculative runtime + optimizations  #496, which contains the full stack, luce_q8_memo is still declared before pools and the destructor still leaves the final memo entries alive. The teardown bug therefore remains present.
• A merge-tree check of this PR against the head of feat(ds4): add DSpark speculative runtime + optimizations  #496 completed without conflicts, including the changes both branches make to server/CMakeLists.txt.

The fix is complementary to that DS4 work and does not introduce another backend or duplicate its inference changes.

Validation

• CUDA 12.6, GCC 11, RTX 3090, sm_86:
  • built test_cuda_pool_shutdown with -j4
  • ctest -R '^cuda_pool_shutdown$' --output-on-failure passed
• ROCm 7.2.4, Strix Halo gfx1151:
  • clean build from current main with -j4
  • ctest -R '^cuda_pool_shutdown$' --output-on-failure passed
  • two serial test processes both initialized and shut down successfully; the second HIP initialization remained healthy after the first teardown

The known-aborting baseline was not intentionally rerun on gfx1151 because reproducing that abort can poison KFD and require a machine reboot.