fix(core_helpers): map Anthropic model_context_window_exceeded stop reason to length

Anthropic returns stop_reason="model_context_window_exceeded" when generation
stops because the context window filled up. _FINISH_REASON_MAP had no entry for
it, so map_finish_reason logged a warning and returned "stop" — clients cannot
distinguish a truncated response from a complete one, on both chat completions
and the Responses API.

Map it to "length" (consistent with max_tokens and the existing compaction
entry) and cover it in the Anthropic parametrized finish-reason tests.

Fixes #43012

litellm/litellm_core_utils/core_helpers.py

@@ -199,6 +199,7 @@ _FINISH_REASON_MAP: Final[dict[str, OpenAIChatCompletionFinishReason]] = {
     "stop_sequence": "stop",
     "end_turn": "stop",
     "max_tokens": "length",
+    "model_context_window_exceeded": "length",
     "tool_use": "tool_calls",
     "refusal": "content_filter",
     "compaction": "length",

tests/unit/litellm_core_utils/test_core_helpers.py

@@ -178,6 +178,7 @@ class TestMapFinishReasonAnthropic:
             ("stop_sequence", "stop"),
             ("end_turn", "stop"),
             ("max_tokens", "length"),
+            ("model_context_window_exceeded", "length"),
             ("tool_use", "tool_calls"),
             ("compaction", "length"),
             ("content_filtered", "content_filter"),