Reduce tokens in generated system prompt

[jiahaog]

Motivation: Better inference latency, accuracy

The following code is some minimal code that generates a system prompt:

schema_manager = manager.A2uiSchemaManager(
    version=constants.VERSION_0_9,
    catalogs=[
      # ...
    ],
)

system_prompt = schema_manager.generate_system_prompt(
    role_description="...",
    include_schema=True,
).strip()

print(system_prompt)

A2UI/agent_sdks/python/src/a2ui/core/schema/manager.py

Line 199 in 7321333

def generate_system_prompt(

The generated system prompt has this structure:

<role description>

## Workflow description

The generated response MUST...

---BEGIN A2UI JSON SCHEMA---

### Server to Client Schema:

<server to client schema>

### Common Types Schema:

<common types schema>

### Catalog Schema:

<catalog schema>

---END A2UI JSON SCHEMA---

This task relates to both the <server to client schema> and <common types schema>.

1. <common types schema> can be automatically tree-shaken based on the <catalog schema> by traversing $refs. Recursively look up $refs, then drop unused types from <common types schemna>.
2. <server to client schema> needs to be tree-shaken with intent from the user. Perhaps generate_system_prompt should be parameterized by a list of allowed messages (e.g. CreateSurfaceMessage)?

@nan-yu @jacobsimionato I'm still new to all this but let me know your thoughts!