parakeet-cpp: Android dynamic backend loading + Adreno-tier GPU policy

[GustavoA1604]

Summary

Brings parakeet-cpp's Android backend story up to parity with
qvac/packages/llm-llamacpp:

• Vulkan and OpenCL ship as dlopen'd MODULE .so files (qvac-ext-ggml@speech's
  GGML_BACKEND_DL=ON), discovered at runtime via ggml_backend_load_all_from_path().
• Zero static GPU backend init calls anywhere in libparakeet. Verified on host:
  nm libparakeet.dylib | grep ggml_backend_\(vulkan\|opencl\|metal\|cuda\|blas\)_init
  returns empty.
• Backend selection mirrors llm-llamacpp's
  BackendSelection.cpp
  tier policy: Adreno 700+ → OpenCL, every other GPU → Vulkan (or Metal / CUDA
  on the matching platform).

What's changed

src/parakeet_ctc.cpp — init_gpu_backend

• Rewrote the registry walk to bucket GPU/IGPU devices into
  {opencl_adreno_700plus, other_gpu, opencl_other} and pick per the tier policy,
  replacing the previous "first GPU/IGPU in registry order, skip Adreno 6xx" logic.
• parse_adreno_version() handles the standard "Adreno 7xx/8xx" naming AND the
  Snapdragon X Elite "Adreno X" naming (mapped to synthetic 800 so it takes the
  OpenCL branch). Existing PARAKEET_ALLOW_ADRENO_6XX env override preserved.
• New public entry points set_backends_directory(dir) / set_opencl_cache_dir(dir)
  (declared in parakeet_ctc.h) so embedded host apps can point the ggml-backend
  registry at a custom per-module folder before the first Engine construction.
  Both honour a "first Engine wins" contract gated on a new g_backends_loaded
  atomic flipped under the shared mutex before the load-all call inside
  ensure_backends_loaded releases it — racing setters either land their value
  (and have it picked up by the in-flight load) or atomically observe the flag
  and fall into the warn-once branch.

include/parakeet/engine.h — EngineOptions

• backends_dir — forwarded to ggml_backend_load_all_from_path() on first
  Engine construction. Empty → ggml's compile-time default search path.
• opencl_cache_dir — Android-only, sets $GGML_OPENCL_CACHE_DIR for
  ggml-opencl's program-binary cache (qvac-ext-ggml@speech program-binary cache
  patch). Strongly recommended in production on Android to skip the cold
  clBuildProgram cost.

src/parakeet_engine.cpp

• Engine ctor calls set_backends_directory / set_opencl_cache_dir before
  load_from_gguf when the respective EngineOptions fields are non-empty.

src/main.cpp

• New --backends-dir DIR CLI flag with the same lifetime contract as
  --opencl-cache-dir (applied before any backend init).

CMakeLists.txt

• On Android, defaults GGML_BACKEND_DL=ON + GGML_CPU_ALL_VARIANTS=ON +
  GGML_CPU_REPACK=ON + GGML_VULKAN=ON + GGML_OPENCL=ON +
  GGML_VULKAN_DISABLE_COOPMAT{,2}=ON, matching the qvac llm-llamacpp Android
  port (qvac-registry-vcpkg/ports/llama-cpp/portfile.cmake).
  Override at the cmake command line as usual.
• Fixed the PARAKEET_GGML_LIB_PREFIX block: it now sets
  GGML_LIB_OUTPUT_PREFIX="speech-" as a cache variable before
  add_subdirectory(ggml), and the post-hoc rename loop is removed. The previous
  version would double-rename when consumed by qvac-ext-ggml@speech's default
  prefix qvac-speech-, producing libspeech-qvac-speech-ggml-vulkan.so style
  filenames that nothing on the runtime side discovered.
• Dropped the dead GGML_USE_VULKAN / GGML_USE_OPENCL / GGML_USE_METAL /
  GGML_USE_CUDA / GGML_USE_BLAS defines from parakeet-backend-defs and the
  parakeet_apply_backend_defs() helper. No source in parakeet-cpp uses
  #ifdef GGML_USE_* anymore (everything goes through the registry); shipping
  these defines would falsely advertise a static backend dependency that the
  GGML_BACKEND_DL=ON Android/Linux builds explicitly do not have.

scripts/setup-ggml.sh

• Bumped to point at qvac-ext-ggml@speech (which carries the speech-stack patch
  series + the qvac-speech- lib filename prefix this PR's prefix change relies
  on).

Companion change

This PR depends on a small ggml-backend loader patch landing on
qvac-ext-ggml@speech: an Android __ANDROID__ block in
ggml_backend_load_best that enumerates per-arch CPU variant names
(cpu-android_armv{8.0,8.2,8.6,9.0,9.2}_*) as candidates for the bare-name
dlopen fallback. Without it, GGML_CPU_ALL_VARIANTS=ON builds on Android
fail to register the CPU backend at runtime (the APK's compressed .so layout
under useLegacyPackaging=false leaves nothing for fs::directory_iterator
to scan, and the existing fallback only composed the base name
libqvac-speech-ggml-cpu.so — which doesn't exist with CPU_ALL_VARIANTS).
Mirrors the equivalent fallback already present downstream on
qvac-fabric-llm.cpp's ggml fork.

tetherto/qvac-ext-ggml#11

Testing

• Host build: cmake -S . -B build -DPARAKEET_BUILD_EXECUTABLES=ON && cmake --build build configures clean and produces correctly-prefixed
  libspeech-ggml-{base,vulkan,opencl,cpu,blas,metal}.dylib files alongside
  libparakeet.dylib. No static GPU backend symbols leaked into libparakeet
  (verified with nm).
• On-device Android via tetherto/qvac-test-addon-mobile against the
  consumer integration suite, on Samsung S23 FE (Cortex-A78, ARMv8.2 + dotprod +
  i8mm):
  • Engine constructs successfully against a q4_0 GGUF (TDT, EOU,
    Sortformer).
  • The Android per-arch CPU fallback picks libqvac-speech-ggml-cpu-android_armv8.2_2.so
    via the score function — no rc=10 from init_cpu_backend anymore.
  • Tier policy correctly selects Vulkan when useGPU=true and the device
    isn't an Adreno 7xx+.
  • Integration tests (runAccuracyMultilangTest, runMultipleTranscriptionsTest,
    runColdStartTimingTest, runDuplexStreamingTest, runEouStreamingTest,
    runMobilePerf*Test, etc.) now actually exercise the engine instead of
    fail-fast on loadGgufOrSkip.