Chore: clean up LLM (prompt caching, supports fn calling), leftover renames

[enyst]

Give a summary of what the PR does, explaining any non-trivial design decisions

This PR does a few clean-ups/leftovers

• Anthropic prompt caching is no longer in beta, so the header is not needed
• max cache markers is 4, remove potential bug with a 5th
• compute whether native function calling is enabled only once per LLM instance
• refactor out Cloudflare exception from llm.py

Also

• remove leak from runtime
• some leftover renaming eventstream->docker (for the runtime)

---

Link of any specific issues this addresses
Fix #4484
Fix #6085

---

To run this PR locally, use the following command:

docker run -it --rm   -p 3000:3000   -v /var/run/docker.sock:/var/run/docker.sock   --add-host host.docker.internal:host-gateway   -e SANDBOX_RUNTIME_CONTAINER_IMAGE=docker.all-hands.dev/all-hands-ai/runtime:ab57782-nikolaik   --name openhands-app-ab57782   docker.all-hands.dev/all-hands-ai/openhands:ab57782

[enyst]

The diff in llm.py looks like a lot of change, but it's not: most of it is only indentation, because we are cleaning out this try / except.

[enyst]

This is tested in CI and it works just fine, thanks to the compatibility code here it's doing the same thing. This PR just names more things "docker runtime".

[enyst]

Prompt caching works without the headers now: (Claude Sonnet 3.5)

17:40:27 - openhands:DEBUG: llm.py:530 - Cost: 0.00 USD | Accumulated Cost: 0.03 USD
Response Latency: 2.561 seconds
Input tokens: 4175 | Output tokens: 96
Input tokens (cache hit): 4033
Input tokens (cache write): 139

[xingyaoww]

LGTM! Thanks for cleaning this up!