fix(py): fix various IDE type check issues, streaming syntax, and missing default value runtime crash

[huangjeff5]

I noticed these issues while spinning up a new sample project and coding directly in IDE:

fromgenkit.plugins.google_genaiimportGoogleAI# ^ red squiggly: "GoogleAI is not exported"ai=Genkit(plugins=[GoogleAI()])
@ai.flow()asyncdefgreet(name: str='world') ->str:
returnf'hi {name}'awaitgreet('Alice') # red squiggly: "str not assignable to None"awaitgreet() # runs but crashes: INVALID_ARGUMENT@ai.flow()asyncdefstream(topic: str, ctx: ActionRunContext) ->str:
asyncforchunkinai.generate_stream(prompt=topic):
# ^ red squiggly: "not iterable"ctx.send_chunk(chunk.text)
return''

1. Plugin imports flagged as private — the plugin's __all__ was built dynamically: __all__ = [GoogleAI.__name__, VertexAI.__name__, ...]. Pyright only treats a symbol as part of the public API when __all__ contains string literals. Rewrote the lists to string literals across google-genai, ollama, and flask __init__.py.

2. Defaulted flow input collapses to None — the @ai.flow() decorator's @overload definitions were ordered most-general first (0-arg → 1-arg → 2-arg). A function async def f(x: str = '...') is structurally callable with no args, so pyright matched the 0-arg overload, which says "input type is None". Reordered to most-specific-first so the input type stays as the caller wrote it.

3. Streaming reported as not iterable — ModelStreamResponse only exposed .stream and .response. Added __aiter__ delegating to the underlying channel. The explicit .stream / .response API still works; the natural async for chunk in ai.generate_stream(...) form now works too.

4. Runtime crash on await greet():
   (A) Action.run() validated input against the schema before checking whether the function had a Python default, so it raised INVALID_ARGUMENT on None against str.
   (B) Even bypassing validation, _invoke was calling self._fn(None), but Python only uses a parameter's default when the argument is omitted, not when None is passed explicitly.

Two-part fix:

• At decoration time, inspect the function's FullArgSpec and remember whether the input arg has a Python default.
• At runtime, skip schema validation when input is missing AND the function has a default, AND omit the input argument from the dispatch call (passing ctx by keyword in the 2-arg form so it doesn't accidentally land in the input slot).

[gemini-code-assist]

Code Review

This pull request enhances the handling of actions with defaulted input arguments, enabling them to be called without input while skipping validation. It also improves static analysis by reordering decorator overloads and transitioning __all__ exports to string literals. A bug was identified in the logic for detecting optional arguments in methods, where the self parameter prevents the correct identification of defaulted arguments.