Allow embedding ced.cpp in another ggml project

[localai-org-maint-bot]

What changed

• CMakeLists.txt: guards the GGML_* cache variable forwarding, the
  GGML_NATIVE/GGML_LLAMAFILE defaults, and add_subdirectory(third_party/ggml)
  behind a CED_TOP_LEVEL check (is ced.cpp the top-level CMake project?) and
  a NOT TARGET ggml check. When ced.cpp is added with add_subdirectory by a
  host project that already configured its own ggml target, ced reuses it
  instead of configuring a second one.
• CMakeLists.txt / src/audio_io.cpp: new CED_EXTERNAL_DR_WAV option
  (default OFF). When ON, ced does not define DR_WAV_IMPLEMENTATION, so the
  host project can provide the one dr_wav implementation in the final binary
  instead of linking two and hitting multiple definition of drwav_*.
• tests/CMakeLists.txt: fixture and include paths now resolve from
  PROJECT_SOURCE_DIR (ced's own project root) instead of CMAKE_SOURCE_DIR
  (the outermost project root), so the ctest targets still find
  models/ and tests/fixtures/ when ced is not the top-level project.
• include/ced_capi.h / src/ced_capi.cpp: new
  ced_capi_classify_pcm_probs(ctx, samples, n_samples, sample_rate, out, n_out),
  which writes every class score in class-index order (no sorting, no
  allocation) and returns the count written, or -1 on error. Callers that
  want the raw per-class distribution on every window (rather than a sorted
  top-k) no longer have to reimplement do_classify + memcpy themselves.
• tests/test_capi.cpp: covers the new function against the existing
  sorted top-k path (same scores at the same indices), a short output buffer
  (n_out < num_classes), and the null-argument error paths.
• README.md: documents CED_EXTERNAL_DR_WAV, the ggml-reuse behavior, and
  ced_capi_classify_pcm_probs in the C-API section.

Why

parakeet.cpp is adding ced.cpp as a git submodule for scene-sound tagging.
parakeet.cpp already vendors its own ggml and its own dr_wav, so ced.cpp
needs to build inside that host project without configuring a second ggml
or linking a second dr_wav implementation, and parakeet.cpp's streaming
scene detector needs the full unsorted score vector per window rather than
a sorted top-k.

How to verify

Standalone build is unaffected (Step 7 in the task brief):

rm -rf build && cmake -B build -DCED_BUILD_TESTS=ON && cmake --build build -j 16
ctest --test-dir build
# 100% tests passed, 0 tests failed out of 7

Embedding in a host project that already has a ggml target and provides
its own dr_wav (Step 8 in the task brief):

mkdir -p /tmp/embed && cd /tmp/embed
ln -sfn <path-to-ced.cpp> ced
cat > CMakeLists.txt <<'EOF'
cmake_minimum_required(VERSION 3.18)
project(embed CXX C)
add_subdirectory(ced/third_party/ggml)        # host project's ggml comes first
set(CED_BUILD_CLI OFF CACHE BOOL "" FORCE)
set(CED_EXTERNAL_DR_WAV ON CACHE BOOL "" FORCE)
add_subdirectory(ced)
add_executable(embed main.cpp)
target_link_libraries(embed PRIVATE ced)
EOF
cat > main.cpp <<'EOF'
#define DR_WAV_IMPLEMENTATION
#include "ced/third_party/dr_wav.h"
#include "ced_capi.h"
int main() { return ced_capi_abi_version() == 1 ? 0 : 1; }
EOF
cmake -B b && cmake --build b -j 16 && ./b/embed && echo EMBED_OK

Builds clean with no multiple definition of drwav_* link errors, ggml is
configured once (by the host project, reused by ced), and the binary prints
EMBED_OK.

This is stacked on #2 (the GPU backend): feat/embedding branches from
feat/gpu-backend, so this diff (against main) also carries #2's commits
until #2 merges.

🤖 Generated with Claude Code

[localai-org-maint-bot]

Landed on main as 61dec2a (squashed; same change as this branch, verified with a diff). Closing.