Fix OpenAI content encoding for thinking parts and empty arrays

[laudney]

Summary

• Strip :thinking ContentParts in the OpenAI encoder — reasoning models (e.g. gpt-oss on vLLM) emit chain-of-thought thinking parts that have no OpenAI encoding. The catch-all returns nil, but an explicit clause makes intent clear.
• Collapse empty content arrays to "" — when all parts are filtered (e.g. assistant message with only :thinking before tool calls), content: [] is rejected by vLLM's strict Pydantic validation. Now normalizes to "" instead.

Test plan

• Tested with gpt-oss-12b on vLLM — multi-turn tool calling with thinking content now works
• Tested with Qwen3-Coder on vLLM — no regression
• Add unit test for :thinking content part encoding
• Add unit test for empty content array normalization

[laudney]

Hey! I've added the missing unit tests that were flagged in the test plan:

• :thinking content part encoding — verifies thinking parts are stripped and remaining text flattens correctly
• Empty content array normalization — verifies content: [] collapses to "" (the vLLM/strict Pydantic fix)
• Also updated the existing "encodes tool calls correctly" test expectation from content: [] to content: "" to align with the fix

All 32 tests in defaults_test.exs pass. Would appreciate another look when you get a chance — thanks!

[laudney]

Rebased onto upstream main — merge conflict resolved.