[FEATURE] Expose `system_prompt_content` as a public property on `Agent`

[lizradway]

Problem Statement

Agent internally maintains two representations of the system prompt:

• self._system_prompt (string) — exposed publicly via the system_prompt property
• self._system_prompt_content (list of SystemContentBlock) — private, no public accessor

The content block representation is the richer, more complete one — it preserves cache control markers and non-text content blocks. The string is a lossy flattened projection kept for backwards compatibility.

Despite being private, _system_prompt_content is already accessed from multiple places outside agent.py:

• event_loop.py passes agent._system_prompt_content to stream_messages()
• Model tests (test_bedrock.py, test_openai.py, test_litellm.py) test system_prompt_content formatting
• test_agent_skills.py directly reads/writes agent._system_prompt_content ~12 times to set up test state
• PR feat(telemetry): emit system prompt on chat spans per GenAI semconv #1818 proposes passing agent._system_prompt_content to telemetry spans

This is a leaky abstraction — the private attribute is effectively part of the public contract already.

Proposed Solution

Add a public system_prompt_content property:

@property
def system_prompt_content(self) -> list[SystemContentBlock] | None:
    """Get the system prompt as content blocks.

    Returns the full content block representation, preserving cache control
    and non-text content. Returns None if no system prompt is set.
    """
    return self._system_prompt_content

Then update internal consumers (event_loop.py, tests) to use the public property instead of the private attribute.

Use Case

The current system_prompt property returns str | None, which was the original API. When content block support was added (for features like prompt caching), _system_prompt_content was introduced as an internal detail. But the string-only public API is now insufficient for consumers that need the full representation — telemetry (#1818), the event loop, and model implementations all need the content blocks.

Alternatives Solutions

No response

Additional Context

No response