whisper : NULL-guard CPU device lookup in make_buft_list (fix #3497)

Three call sites in `src/whisper.cpp` dereference the CPU backend device
returned by GGML without checking for NULL. With GGML_BACKEND_DL=ON
builds, the registry starts empty and an embedder that calls
`whisper_init_*` or `whisper_vad_init_*` before any explicit
`ggml_backend_load*()` invocation hits `GGML_ASSERT(device != NULL)`
inside `ggml_backend_dev_backend_reg` and the process is killed with
`__fastfail(7)` / `0xC0000409` on Windows.

Reported in #3497 (Flutter Windows embedder, same symptom). Independently
reproduced on a Windows GGML_BACKEND_DL=ON build that calls
`whisper_vad_init_from_file_with_params` before the dynamic CPU backend
DLL is registered.

Sites patched (line numbers from current master):

- `make_buft_list` (line 1388): `cpu_dev` from `ggml_backend_dev_by_type`
  can be NULL. Return an empty buft list so the caller surfaces a NULL
  context instead of aborting.
- `ggml_graph_compute_helper` (line 173): `backend` from
  `ggml_backend_init_by_type(CPU)` can be NULL. Return false so the
  caller surfaces a typed graph compute failure.
- `whisper_exp_compute_token_level_timestamps_dtw` (line 8960): same
  pattern; skip the DTW step for this segment rather than abort.

All three log a clear WHISPER_LOG_ERROR message pointing embedders at
`ggml_backend_load_all()` / `ggml_backend_load()` as the fix on their
side. Mirrors the existing NULL-guard upstream already has at line
1352 in `whisper_backend_init`.

Build verified locally on macOS (CMake configure + build of the
`whisper` shared library target, 0 warnings).

src/whisper.cpp

@@ -171,6 +171,13 @@ static bool ggml_graph_compute_helper(
          ggml_abort_callback   abort_callback,
                         void * abort_callback_data) {
     ggml_backend_ptr backend { ggml_backend_init_by_type(GGML_BACKEND_DEVICE_TYPE_CPU, nullptr) };
+    if (!backend) {
+        // No CPU backend registered (GGML_BACKEND_DL=ON without a prior
+        // ggml_backend_load_all() call). Avoid the NULL deref a few lines
+        // below and surface the failure instead of aborting.
+        WHISPER_LOG_ERROR("%s: failed to initialize CPU backend\n", __func__);
+        return false;
+    }
 
     auto * reg = ggml_backend_dev_backend_reg(ggml_backend_get_device(backend.get()));
 
@@ -1386,6 +1393,20 @@ static buft_list_t make_buft_list(whisper_context_params & params) {
 
     // CPU Extra
     auto * cpu_dev = ggml_backend_dev_by_type(GGML_BACKEND_DEVICE_TYPE_CPU);
+    if (cpu_dev == nullptr) {
+        // No CPU backend registered in the GGML registry. This can happen with
+        // GGML_BACKEND_DL=ON builds when the host process has not loaded any
+        // CPU backend DLL via `ggml_backend_load*()` before reaching this code
+        // path (e.g. `whisper_init_*` or `whisper_vad_init_*` called before any
+        // explicit backend load). Without this guard, the next line aborts
+        // the process via `GGML_ASSERT(device != NULL)`. Returning an empty
+        // buft_list lets the caller surface a NULL whisper context instead.
+        WHISPER_LOG_ERROR("%s: no CPU backend device registered; "
+            "load a CPU backend with ggml_backend_load_all() or "
+            "ggml_backend_load() before initializing the whisper context\n",
+            __func__);
+        return {};
+    }
     auto * cpu_reg = ggml_backend_dev_backend_reg(cpu_dev);
     auto get_extra_bufts_fn = (ggml_backend_dev_get_extra_bufts_t)
         ggml_backend_reg_get_proc_address(cpu_reg, "ggml_backend_dev_get_extra_bufts");
@@ -8937,6 +8958,13 @@ static void whisper_exp_compute_token_level_timestamps_dtw(
     ggml_build_forward_expand(gf, w);
 
     ggml_backend_ptr backend { ggml_backend_init_by_type(GGML_BACKEND_DEVICE_TYPE_CPU, nullptr) };
+    if (!backend) {
+        // No CPU backend registered (GGML_BACKEND_DL=ON without a prior
+        // ggml_backend_load_all() call). Skip DTW alignment for this
+        // segment instead of aborting the whole process.
+        WHISPER_LOG_ERROR("%s: failed to initialize CPU backend for DTW; skipping token-level timestamps\n", __func__);
+        return;
+    }
     ggml_backend_graph_compute(backend.get(), gf);
 
     ggml_tensor * alignment = dtw_and_backtrace(gctx, w);