feat: expose current session to model plugins

[ryuta-kobayashi-ug]

Fixes #311

Follow-up to langgenius/dify#35772.

Summary

Expose the current SDK Session to model plugins through get_current_session().
This lets provider plugins read session.app_id for provider-side request metadata and cost attribution.

Current host-side design

The earlier plan changed substantially.
The host side now passes app identity explicitly before the request reaches this SDK:

• feat(model-runtime): thread request_metadata through to model runtimes graphon#191 forwards keyword-only request_metadata through model runtime protocols.
• feat(api): pass app_id to model plugins for provider-side cost attribution dify#35859 is merged and passes request_metadata["app_id"] into plugin model dispatch payloads.
• The plugin daemon already maps that payload into the SDK Session.

That leaves one SDK problem: model plugin implementations do not receive Session.
Tool plugins receive it in their constructor, but model plugin _invoke() signatures are implemented by existing third-party plugins.
Adding a new _invoke() parameter would break those plugins.
AIModel.__init__() is final, so constructor injection is not the right extension point either.

Changes

• Add dify_plugin.core.session_context.get_current_session().
• Export get_current_session from the top-level dify_plugin package.
• Bind the request Session around model invocation paths in PluginExecutor.
• Wrap LLM streaming generators so the session remains available until the stream is consumed.
• Add focused tests for the public API, synchronous model invocation, and streaming LLM invocation.

Usage

from dify_plugin import get_current_session

session = get_current_session()
if session and session.app_id:
    # attach session.app_id to provider request metadata
    ...

session.app_id is None for model calls outside app execution, such as routing, title generation, suggested questions, or other system use.
Provider plugins should skip tagging or bucket those calls as internal/system traffic.

Validation

• just check
• just test

[gemini-code-assist]

Code Review

This pull request implements request-scoped session context propagation for model plugins using Python's contextvars. It introduces a get_current_session utility and updates the PluginExecutor to manage session state across various model invocation types. A critical issue was identified in the invoke_llm implementation where managing the ContextVar outside of the generator's execution flow could lead to context leaks or ValueError exceptions. The reviewer suggests moving the context management logic entirely inside the generator wrapper to ensure proper cleanup.