whisper : load backends before loading the model

This commit adds a call to `whisper_load_backends()` in the
`whisper_model_load` function.

The motivation for this change is that currently is `GGML_BACKEND_DL` if
enabled the model loading will crash because of a segment fault as the
`cpu_dev` will be null in `make_buft_list` becasue it has not been loaded.

There have been other suggestions to address this issue that were more
involved than this change so there is a good chance that I missing
something here, but this seemed like a simple change to make.

Refs: #3182

Author: danbev

src/whisper.cpp

@@ -1506,6 +1506,7 @@ static ggml_backend_buffer_type_t select_weight_buft(const whisper_hparams & hpa
 //
 static bool whisper_model_load(struct whisper_model_loader * loader, whisper_context & wctx) {
     WHISPER_LOG_INFO("%s: loading model\n", __func__);
+    whisper_load_backends();
 
     const int64_t t_start_us = ggml_time_us();