Add tests for consistent token counts in streaming vs non-streaming responses

- Implemented a new test in `chat_streaming_spec.rb` to verify that token counts for streaming responses are consistent with non-streaming responses for various models, excluding the DeepSeek provider due to known discrepancies.
- Removed specific skips for the `claude-3-5-haiku-20241022` model in `chat_tools_spec.rb` to allow for more robust testing across different models.
- Added VCR cassettes for streaming and non-streaming responses for multiple models including Anthropic Claude, Bedrock, Gemini, and OpenAI GPT-4.1 to facilitate consistent testing.

Author: crmne

lib/ruby_llm/stream_accumulator.rb

@@ -89,8 +89,8 @@ def find_tool_call(tool_call_id)
     end
 
     def count_tokens(chunk)
-      @input_tokens += chunk.input_tokens if chunk.input_tokens
-      @output_tokens += chunk.output_tokens if chunk.output_tokens
+      @input_tokens = chunk.input_tokens if chunk.input_tokens
+      @output_tokens = chunk.output_tokens if chunk.output_tokens
     end
   end
 end