refactor(workflow): move output validation off inline node paths

[wolo-lab]

Problem

Output validation for FunctionNode and ToolNode was duplicated. Each node
validated its output inline (inside wrappedFn / runTool) and the
scheduler validated it again scheduler-side via BaseNode.ValidateOutput
(introduced in #970). The same OutputSchema was therefore checked twice on
every activation.
Beyond the redundancy, the two copies are a drift hazard: as the
scheduler-side helper evolves (e.g. #970 added the *genai.Content text
fallback and passthrough for framework control values), the inline copies fall
behind and the two paths diverge. The validation behaviour also stopped being
uniform across node types — custom Node implementations only ever went through
the scheduler, while FunctionNode/ToolNode carried their own logic.
The FunctionTool {"result": X} unwrap convention was buried inline inside
ToolNode.runTool, mixing a tool-specific quirk into the execution path instead
of expressing it as part of the validation contract.

Solution

Make the scheduler the single, authoritative output-validation point and remove
the inline validation from the node run paths:

• FunctionNode: drop the inline OutputSchema check from all three
  constructors (newFunctionNodeWithResolvedSchemas, NewFunctionNodeFromState,
  and the emitting variant). The schema still flows through BaseNode, so the
  scheduler enforces it via BaseNode.ValidateOutput. wrappedFn is back to
  doing one thing: materialise the typed parameter and call the user function.
• ToolNode: remove the inline validation from runTool and move the
  FunctionTool {"result": X} unwrap into a dedicated ToolNode.ValidateOutput
  override. The override tries standard validation first, falls back to
  unwrapping {"result": X}, and re-runs standard validation on total failure
  so the caller still sees the original schema-mismatch error. This is the right
  home for the convention because it is tool-specific — making it a general
  default could mask genuine validation errors in other node types.

Behaviour note

Validation errors now surface scheduler-side rather than inside Run. For real
workflows this is unobservable (every activation goes through the scheduler).
Callers that invoke node.Run() directly (mostly tests) now receive the raw
output and must call ValidateOutput themselves. Tests are split accordingly:
Run-level tests assert raw passthrough; validation is asserted via
ValidateOutput (TestFunctionNode_ValidateOutput,
TestToolNode_ValidateOutput covering direct-valid, {"result": X} unwrap,
fail-both-paths, and nil-schema cases).

[hanorik]

Maybe to add a success test case as well?

[wolo-lab]

Good idea. Adding.