[Observability] Clean up input and output messages from auto-instrumented invoke_agent spans

[threddy]

Why?
Existing clean up logic does not work as expected because:
It attempts to cleanup invocation_input and invocation_output attributes, but for invoke_agent spans SK uses input_messages and output_messages
Deserialization fails because it expects a list of strings but actually it is a list of objects of type MessageContent.

What?
Try to filter input_messages and output_messages attributes in addition to existing.
First try to deserialize into a list of objects.

Testing

Without fix

Zooming in on input_messages and output_messages

With fix

[Copilot]

Pull request overview

This PR fixes a bug in the observability cleanup logic for auto-instrumented spans from Semantic Kernel. The existing cleanup attempted to filter invocation_input and invocation_output attributes but failed because: (1) SK actually uses input_messages and output_messages attributes for invoke_agent spans, and (2) deserialization expected a list of strings but received a list of MessageContent objects.

Key Changes:

• Enhanced deserialization to handle JSON arrays of MessageContent objects directly, with fallback to string array format
• Added filtering support for gen_ai.input.messages and gen_ai.output.messages attributes
• Refactored code to reduce duplication by introducing a helper overload

Reviewed changes

Copilot reviewed 2 out of 2 changed files in this pull request and generated 2 comments.

File	Description
src/Observability/Extensions/SemanticKernel/Utils/SemanticKernelSpanProcessorHelper.cs	Refactored filtering logic to support additional message attribute keys (input_messages, output_messages) and fixed deserialization to handle JSON arrays of objects directly
src/Tests/Microsoft.Agents.A365.Observability.Extension.Tests/SemanticKernelSpanProcessorHelperTests.cs	Added test verifying the new deserialization logic correctly handles JSON arrays of MessageContent objects

[nikhilNava]

The ask from the agents activity page team is they want human readable text.
From the images we only require value of the content field e.g "Hello! Before I can ....."

[threddy]

The ask from the agents activity page team is they want human readable text. From the images we only require value of the content field e.g "Hello! Before I can ....."

Updated to handle nested content. See new "after fix" screenshot in PR descritpion