docs: clarify agent runtime, handoff, HITL, and results flows

[seratch]

This pull request updates the Python documentation to make the SDK's runtime surfaces easier to navigate and reason about across multi-agent workflows. It rewrites the result guide around surface selection instead of a flat property inventory, clarifies handoff input typing and guardrail boundaries, expands the HITL guide around run-wide approvals and resumable flows, and tightens the docs for context, sessions, and model-input shaping so the guides agree on how state moves through a run.

The main changes are in docs/results.md, docs/human_in_the_loop.md, docs/handoffs.md, docs/context.md, and docs/agents.md, with supporting clarifications in docs/guardrails.md, docs/running_agents.md, docs/sessions/index.md, and the generated handoff reference docstrings in src/agents/handoffs/__init__.py.

Behavioral/documentation clarifications included in this PR:

• handoff(input_type=...) is documented as the handoff tool-call payload schema rather than the next agent's main input.
• HITL approvals are documented as run-wide across direct tools, handoffs, and nested Agent.as_tool() runs, including partial resolution and resume behavior.
• Result surfaces now explain the relationship between input, to_input_list(), new_items, raw_responses, last_response_id, agent_tool_invocation, and streaming lifecycle caveats.
• Context and session docs now distinguish local app context from conversation state, explain what RunContextWrapper exposes, and clarify when to use session_input_callback versus call_model_input_filter.
• The agents guide now acts as a navigation hub for related runtime guides and calls out key agent configuration fields more explicitly.