Describe the bug

The Anthropic prompt caching implementation had two issues that affected cache efficiency:

1. Cache prefix continuity broken: MOIM (Message of Instruction Mode) messages were being inserted after the last assistant message, which broke the cache prefix continuity required by Anthropic's caching system.

2. Cache control logic complexity: The cache_control logic used nested loops and didn't properly exclude MOIM messages (<info-msg>) from caching, potentially caching invalid content types.

To Reproduce

Steps to reproduce the behavior:

1. Use goose with Anthropic provider (Claude models)
2. Have a conversation with multiple turns
3. Observe MOIM insertion point in conversation
4. Check that cache_control is applied to MOIM messages when it shouldn't be
5. Notice reduced cache hit rates due to broken cache prefix

Expected behavior

• MOIM messages should be inserted at the end of the conversation to preserve Anthropic's cache prefix continuity
• Only valid content block types (text, image, tool_use, tool_result) should have cache_control applied
• MOIM messages (<info-msg>) should be excluded from caching

Please provide the following information

• OS & Arch: [All platforms]
• Interface: [UI / CLI]
• Version: [Affected versions prior to commit 64c4bc1]
• Extensions enabled: [All]
• Provider & Model: [Anthropic – Claude models with prompt caching enabled]

Additional context

This fix improves:

• Cache hit rates for Anthropic conversations
• Cost efficiency when using prompt caching
• Code maintainability with simplified cache control logic

The issue affected all users of goose with Anthropic provider when prompt caching was enabled.