Implement A2A extension activation via A2A-Extensions header

[tomkis]

Is your feature request related to a problem? Please describe.
The A2A spec requires extensions to be activated per-request via the A2A-Extensions HTTP header (spec). Extensions should default to inactive. Currently neither the server nor the client implements this.

Server (agentstack-sdk-py):

• Extension dependencies are always injected regardless of client activation
• Outgoing message metadata (citations, trajectories, etc.) is never filtered based on activated extensions
• No A2A-Extensions response header is set

Client (agentstack-sdk-ts / agentstack-ui):

• handleAgentCard reads agent card extensions to resolve service demands (LLM, MCP, etc.) via message metadata, but never sends A2A-Extensions header
• UI extensions (citation, trajectory, error, canvas) are parsed from response metadata unconditionally — client never declares which UI extensions it actually supports
• sendMessageStream call has no mechanism to attach the activation header

Describe the solution you'd like

• Client should send A2A-Extensions header listing URIs of extensions it supports
• Server should read the header and only include metadata for activated extensions in responses
• Server should set A2A-Extensions response header confirming activated extensions
• BaseExtensionServer should expose activation state so agent code can branch on it

Additional context
May require upstream changes in a2a-sdk (Python) and @a2a-js/sdk if request/response contexts don't carry header info today.

[jezekra1]

Slight improvement proposal

When refactoring extensions, consider explicitly marking them as required or not required, e.g. to avoid lazy implementations with faux fallbacksm e.g.:

if not llm_ext:
  yield "agent needs llm to work"

Proposal - use "required" and "default" fields in extension spec, for example:

A required extension
Agent won't work without it (needs a specialized client), but at least we are honest about it:

llm_ext: Annotated[
        LLMServiceExtensionServer,
        LLMServiceExtensionSpec.single_demand(required=True),
]

Explicit fallback
If the extension defined but not required, we will force users to provide fallback settings. Of course they can still use "dummy", but it's more guided than the if not llm_ext above + we can make the type non-nullable, which makes the code more type-safe and removes the need for guards like if not llm_ext.

llm_ext: Annotated[
        LLMServiceExtensionServer,
        LLMServiceExtensionSpec.single_demand(
           required=False,
           default=LLMFulfillment(api_key=os.getenv("OPENAI_API_KEY"))
        ),
    ],