Add a handoffs/message_filter sample, fix passing parameters to activity tools, fix ActivityModelInput serialization

[AnatoliB]

Ported handoffs/message_filter sample from OpenAI Agents SDK:

• The tool function returning a random number had to be converted to an activity tool to preserve determinism
• The tool activity definition moved to function_app.py

Fixed two issues exposed by this sample:

• When a list is passed to the input parameter of Runner.run_sync, the ActivityModelInput object failed to serialize to JSON
• Activity tool parameters were passed wrapped to a JSON like this: {"max":100} - fixed by parsing this and extracting the value properly

[philliphoff]

This is fine, but I think there's another serializer function used elsewhere; I wonder if we can combine them at some point.

Also, I wonder if we should add a layer of indirection in the output of the activity rather than return OpenAI types directly, just in case we need to make versioning adjustments in the future--else we may be subject to breaking OpenAI changes. (I had a change in progress that did this.)

[AnatoliB]

This is fine, but I think there's another serializer function used elsewhere; I wonder if we can combine them at some point.

Sure, let's do it later.

Also, I wonder if we should add a layer of indirection in the output of the activity rather than return OpenAI types directly, just in case we need to make versioning adjustments in the future--else we may be subject to breaking OpenAI changes. (I had a change in progress that did this.)

Are you talking about a situation when activity output is saved in the orchestration history, then the user upgrades their app to a new OpenAI SDK version, and they expect this orchestration to continue after an upgrade? Yes, we should be able to eventually handle that, good catch.

[philliphoff]

Yes, that's what I was referring to. Though, thinking about it again, I suppose it doesn't really matter whether the OpenAI type is returned directly or wrapped in an outer type (e.g. ActivityModelOutput). We can still add logic to the post-activity-call handler to make any adjustments needed before passing it back to the OpenAI SDK layer.