Decouple AgentServer responses from generated model internals

[Shivakishore14]

Summary

• localize AgentServer response contracts as TypedDict/dict wire payloads
• remove runtime dependency on generated model constructors, serialization helpers, and nested generated SDK model paths
• regenerate the local response TypedDict models from the source-built Azure/typespec-azure Python emitter that supports models-mode=typeddict
• update storage, streaming, validation, samples, and tests to use dict-native response payloads

Validation

• make generate-models TYPESPEC_PYTHON_EMITTER_PATH=/tmp/typespec-azure-agentserver/packages/typespec-python
• make generate-openapi && make generate-validators
• PYTHONPATH=. /tmp/azsdk-agentserver-cleanup-venv/bin/python -m pytest tests/unit tests/contract tests/interop -q
  • 951 passed, 1 warning

Current source-emitter gaps

• Enum generation is incomplete: the source emitter outputs Literal[...] aliases, but generated annotations still reference EnumName.MEMBER, and the expected _enums.py module is not emitted. The finalizer currently adds enum-like fallback exports and rewrites generated enum-member references to literal strings.
• Generated package is nested too deeply: source output lands under _generated/agentserver/responses/models/; the finalizer extracts the emitted TypedDict files into the local _generated/ package boundary.
• TypedDict mode still emits client/operation files: the source emitter emits _client.py, _configuration.py, operations/, and aio/; the finalizer copies only the model contract files needed by AgentServer.
• Synced TypeSpec overlay has stale OpenAI references: client.tsp still references removed OpenAI symbols (OutputItemComputerToolCallOutputResource, OutputItemCustomToolCall, OutputItemCustomToolCallOutput, FunctionToolCallOutputResource), so the finalizer removes those stale @@usage lines before compilation.
• No published npm package supports this yet: npm latest and next still reject models-mode=typeddict; generation currently requires a source-built emitter from Azure/typespec-azure/main via TYPESPEC_PYTHON_EMITTER_PATH.
• Post-processing is still required: finalization currently patches stale TypeSpec usage refs, extracts nested model output, patches enum references, strips trailing whitespace, and keeps validators generated from OpenAPI.

Upstream asks

• Fix TypedDict enum emission so generated code imports and supports runtime constants without a local fallback.
• Support model-only / flat TypedDict output for this local contract package shape.
• Publish an emitter version with models-mode=typeddict support.
• Update the AgentServer TypeSpec overlay to remove stale OpenAI symbol references.

Notes

• This PR is scoped to sdk/agentserver/azure-ai-agentserver-responses.