Fix Coverity static analysis defects (Feb) (openvinotoolkit#3376)

## Description

Fix three Coverity static analysis defects across speculative decoding,
LLM pipeline, and Whisper components.

- **CID 6415047 (INTEGER_OVERFLOW)**: Add boundary guards before
decrementing unsigned indices in DTW backtracking to prevent `size_t`
underflow -
[CVS-181400](https://jira.devtools.intel.com/projects/CVS/issues/CVS-181400)
- **CID 6419148 (MISSING_LOCK)**: Access `m_draft_generations_mutex`
field directly instead of through accessor in `generate_common`
template, so Coverity can correlate the lock with the guarded data -
[CVS-181402](https://jira.devtools.intel.com/projects/CVS/issues/CVS-181402)
- **CID 6415437 (NULL_FIELD)**: Add defensive assertion on
`main_model_descr.model` before passing it to `StatefulLLMPipeline`
constructor -
[CVS-181401](https://jira.devtools.intel.com/projects/CVS/issues/CVS-181401)

Author: sgonorov

src/cpp/src/llm/pipeline.cpp

@@ -151,8 +151,9 @@ static std::unique_ptr<LLMPipelineImplBase> create(const std::shared_ptr<ov::Mod
 
     auto main_model_descr =
         ov::genai::ModelDesc(model, tokenizer, device, properties_without_draft_model, {}, generation_config);
+    OPENVINO_ASSERT(main_model_descr.model, "Model descriptor must contain a valid model");
 
-    if (draft_model_descr.model != nullptr) {
+    if (draft_model_descr.model) {
         // FIXME: Add support for StatefulSpeculativeLLMPipeline for non-NPU devices for both models.
         OPENVINO_ASSERT(device == "NPU" || draft_model_descr.device == "NPU",
                         "Stateful FastDraft and Stateful Eagle3 Speculative Decoding require NPU to be "

src/cpp/src/speculative_decoding/continuous_batching/fast_draft_strategy.hpp

@@ -59,7 +59,7 @@ std::vector<EncodedGenerationResult> generate_common(
 
     std::vector<GenerationHandle> main_generations;
     {
-        std::lock_guard<std::mutex> lock(self->draft_generations_mutex());
+        std::lock_guard<std::mutex> lock(self->m_draft_generations_mutex);
         for (size_t rid = 0; rid < input_ids.size(); ++rid) {
             GenerationConfig main_cfg = sampling_params[rid];
             GenerationConfig draft_cfg = main_cfg;
@@ -196,7 +196,6 @@ class ContinuousBatchingPipeline::SpeculativeDecodingImpl : public ContinuousBat
     Tokenizer& tokenizer() { return m_tokenizer; }
     const Tokenizer& tokenizer() const { return m_tokenizer; }
 
-    std::mutex& draft_generations_mutex() { return m_draft_generations_mutex; }
 };
 
 }  // namespace ov::genai

src/cpp/src/whisper/word_level_timestamps.cpp

@@ -269,12 +269,17 @@ std::vector<std::pair<size_t, size_t>> dtw_and_backtrace(const std::vector<std::
         path.push_back({path_i, path_j});
         int t = trace[i][j];
         if (t == 0) {
-            i--;
-            j--;
+            OPENVINO_ASSERT(i > 0 && j > 0, "Invalid DTW trace: diagonal move at matrix boundary");
+            --i;
+            --j;
         } else if (t == 1) {
-            i--;
+            OPENVINO_ASSERT(i > 0, "Invalid DTW trace: up move at top boundary");
+            --i;
         } else if (t == 2) {
-            j--;
+            OPENVINO_ASSERT(j > 0, "Invalid DTW trace: left move at left boundary");
+            --j;
+        } else {
+            OPENVINO_THROW("Invalid DTW trace value: ", t);
         }
     }