feat(py): drop Part(root=); one Part through generate

[huangjeff5]

Callers already build with Part.from_text / Part(text=...). This change deletes .root. Part(root=...) raises. The part they put on Message / Document / ai.generate is the same Part they read .text from.

# before
msg = Message(role='user', content=[Part(root=TextPart(text='hi'))])
print(msg.content[0].root.text)

# after
msg = Message(role='user', content=[Part.from_text('hi')])
# or Part(text='hi')
print(msg.content[0].text)

Part.from_data(None) is a data part. Empty {} or metadata-only parts in inbound message content are dropped when the message is read; Part() with no kind still raises.

Part.from_data(None)  # data part, same as 0 or []

# inbound junk is skipped; the rest of the message loads
Message(role='model', content=[{}, {'text': 'hello'}]).text  # 'hello'

Decisions

• Rewrite construction sites to Part.from_* or Part(text=...). Part(root=...) raises.
• Kind classes stay off from genkit import. Inner objects (ToolRequest, ToolResponse, Media) stay public.
• Two exclusive kinds still raise. custom may ride on another kind, or be the kind when it is the only payload.
• JSON null is a real data payload. Part.from_data(None) dumps and reloads as a data part.
• An inbound {} or metadata-only part is dropped when reading a message so a junk empty part does not fail generate. Building Part() with no kind still raises.
• A thought whose summary includes an image replays as that same thought step.
• A typed plugin config on ModelRequest stays that class on the response. /generate still requires messages.
• A bad pendingContent part reports the kind rule (exactly one of text, media, ...) instead of a generic "must be a part".

[gemini-code-assist]

Code Review

This pull request refactors the Part data model across the Genkit Python SDK and its provider plugins (including Bedrock, Anthropic, Google GenAI, Ollama, and OpenAI). It flattens the Part structure, replacing the nested root attribute with direct properties (such as text, media, and tool_request) and introducing helper factory methods like Part.from_text and Part.from_media. The review feedback identifies a critical runtime error where model_dump is called with an unsupported fallback argument on a Pydantic BaseModel. Additionally, several unidiomatic expressions (such as not part.media is not None) and redundant checks can be simplified to improve code readability and maintainability.