[Refactor]: Eliminate wrappedSession and decouple workflow node prompt injection from Session

### Problem Description

In `agent/llmagent/llm_agent_wrapper.go:642-660`, `single_turn` workflow nodes inject transient node inputs into the LLM prompt by wrapping the underlying session in synthetic `wrappedSession` and `wrappedEvents` structs (`PrepareLLMAgentInput`).

This has significant architectural drawbacks (noted by maintainers in `TODO` at `llm_agent_wrapper.go:642`):
1. **Session Pollution**: Wrapping the session pollutes the `session.Session` abstraction: any tool, plugin, callback, or telemetry hook inspecting `ctx.Session()` during execution sees synthetic phantom events that do not exist in persistent storage.
2. **Unnecessary Coupling**: Faking a session event just to feed a prompt to the LLM couples workflow graph node orchestration to session state manipulation.
3. **Index Drift Risk**: `wrappedEvents.At(i)` relies on manual index offset math (`insertAt`), which risks index misalignment during multi-round tool executions.

### Proposed Solution

As outlined in the codebase `TODO`:
1. Eliminate `wrappedSession`, `wrappedEvents`, and `PrepareLLMAgentInput` from `agent/llmagent/llm_agent_wrapper.go`.
2. Rely directly on `InvocationContext.UserContent()` (which already carries the rendered `nodeInput`).
3. Update `internal/llminternal/contents_processor.go` to directly read `ctx.UserContent()` and prepend it to `LLMRequest.Contents` for `single_turn` agents.
4. Keep the session history pure, properly decoupling workflow node inputs from session state.

Read the TODO and `PrepareLLMAgentInput` in `agent/llmagent/llm_agent_wrapper.go:642-660`, then trace `InvocationContext.UserContent()` into `internal/llminternal/contents_processor.go`. The work is done when the wrapped session/event types and helper are removed, single-turn prompt input is prepended from `UserContent()`, and session history remains free of synthetic events.

Written by the indexing model from the issue text.