tts-cpp: Add dynamic backend selection for Android

[GustavoA1604]

What problem does this PR solve?

• Android dynamic loading: Embedded host apps ship ggml backends as separate .so files (libspeech-ggml-vulkan.so, libspeech-ggml-opencl.so, per-arch libspeech-ggml-cpu-android_armv*_*.so) with GGML_BACKEND_DL=ON. The old tts-cpp init path called ggml_backend_{cuda,metal,vk,opencl,cpu}_init() directly — those symbols live in dlopen'd libraries and are not linkable from libtts-cpp.
• Inconsistent stack behavior: parakeet-cpp and llm-llamacpp already use registry-based backend discovery plus an Adreno-tier GPU policy (OpenCL on 7xx/8xx, Vulkan elsewhere, 6xx OpenCL skipped). tts-cpp still used compile-time #ifdef GGML_USE_<X> cascades, so Android prebuilds could silently compile out GPU paths or fail at link/runtime.
• Missing host knobs: No way for QVAC addons to pass backends_dir or an OpenCL program-binary cache directory before the first Engine init — both are process-singleton state in ggml.

How does it solve it?

Registry-only backend selection (src/backend_selection.{h,cpp})

• ensure_backends_loaded() — one-time ggml_backend_load_all() or ggml_backend_load_all_from_path(backends_dir) (idempotent per process).
• init_gpu_backend() — walks ggml_backend_dev_* with the same tier policy as parakeet-cpp:
  1. Adreno 700+ (incl. synthetic 800 for Adreno X<n> / Snapdragon X Elite) → prefer OpenCL
  2. Other GPUs (Vulkan on non-Adreno Android, Metal, CUDA, Mali, …) → non-OpenCL backends
  3. Fallback → other OpenCL devices (desktop, etc.)
  • Adreno 6xx OpenCL is skipped (known broken kernels) unless TTS_CPP_ALLOW_ADRENO_6XX=1.
• init_cpu_backend() / init_blas_backend() — registry helpers for CPU and BLAS accel devices.
• set_backends_directory() / set_opencl_cache_dir() — first-Engine-wins overrides with one-shot stderr warnings on conflict (mirrors parakeet-cpp).

backend_util.h

• backend_is_cpu / backend_is_metal / backend_set_n_threads — route through ggml_backend_get_device + ggml_backend_reg_get_proc_address so they work under GGML_BACKEND_DL=ON (replaces direct ggml_backend_is_cpu / ggml_backend_cpu_set_n_threads calls that fail to link on Android).

Init site refactor

• Replaces duplicated ladders in main.cpp, chatterbox_tts.cpp (s3gen_init_backend), supertonic_gguf.cpp, and related paths with init_gpu_backend + init_cpu_backend.
• Updates auxiliary CPU fallbacks in voice_encoder.cpp, campplus_forward.inc, mel_extract_stft.cpp.

Public API (EngineOptions)

• backends_dir — forwarded to set_backends_directory() before first backend init (Chatterbox + Supertonic engines).
• opencl_cache_dir — sets $GGML_OPENCL_CACHE_DIR on __ANDROID__ before first init (reduces cold clBuildProgram latency).

CMake (tts-cpp/CMakeLists.txt)

• Android defaults (when not already in cache): GGML_BACKEND_DL=ON, GGML_CPU_ALL_VARIANTS=ON, GGML_CPU_REPACK=ON, GGML_VULKAN=ON, GGML_OPENCL=ON, coopmat/coopmat2 disabled — aligned with parakeet-cpp and llamacpp Android prebuilds.
• tts-cpp-backend-defs — kept as an empty INTERFACE target for export compatibility; GGML_USE_* compile defines removed (they falsely implied static backend linkage on Android).
• test-metal-ops — GGML_USE_METAL defined locally on that target only (it exercises Metal directly, not the registry path).

Behavior changes (non-Android)

Desktop / dev builds no longer pick GPUs via compile-order cascade (CUDA → Metal → Vulkan → OpenCL). Selection now follows the Adreno-aware registry policy; on multi-GPU machines the chosen device follows registry enumeration order within each bucket, not #ifdef link order. This matches parakeet-cpp and llamacpp.

Breaking changes

None for API consumers that only pass existing EngineOptions fields. New optional fields (backends_dir, opencl_cache_dir) are additive.

Android prebuilds will emit additional backend .so artifacts (Vulkan, OpenCL, per-arch CPU) that must be staged next to the host module — same layout as parakeet / llamacpp.