Bug: Token usage mismatch between Kernel token usage and custom HttpClient response

[CodePlacer]

we noticed a discrepancy in token usage between the values retrieved using custom http client and the token usage reported in the agent's chat response.

Preproducing Steps:

1. Create a Kernel Object
2. Import Lights plugin to kernel and set ToolCallBehavior to AutoInvoke
3. Prompt to turn on the lamp.
4. Note the token usage from function result.
5. After 4th step, pass custom http client handler to kernel object and parse the response object and track the token usage from response body.

On Analysis it look like , the Token Usage available in kernel object is only showing the usage of last LLM calls. Token usage of tool calls are not captured properly.

Sample code is added to https://github.com/CodePlacer/Semantic-Kernel/tree/main/TokenDiff