Incorrect completion tokens in UsageInfo for Gemini via VertexAI

[sam-veezoo]

For Gemini models accessed via VertexAI, the completion_tokens field of the UsageInfo object seems to contain the number of input tokens, but I think it should be the number of output tokens.

Here is a demonstrating example:

val modelId = "gemini-2.0-flash"

val completionService: OpenAIChatCompletionService =
  VertexAIServiceFactory.asOpenAI(
    projectId = ...,
    location = ...
  )

val response = Await.result(completionService.createChatCompletion(
    Seq(domain.UserMessage("Please output the exact string 'ABC' excluding the quotes and nothing else")),
    CreateChatCompletionSettings(modelId)
  ),
  Duration.Inf
)

val usage = response.usage.get

val inputTokens = usage.prompt_tokens           // returns 14 (OK)
val outputTokens = usage.completion_tokens.get  // returns 14 (EXPECTED: 2)
val totalTokens = usage.total_tokens            // returns 16 (OK)

This seems to be a regression, as this piece of code was returning correct numbers in version 1.1.0.