llama.cpp b10219: reasoning_content persists across chat turns

llama.cpp b10219 fixes a bug where llama-cli discarded reasoning tokens between turns. The --reasoning-preserve flag is supposed to feed prior reasoning back into context on later turns — but until now, it couldn’t, because the reasoning was never stored.

The fix (PR #26362) persists reasoning_content in the message history:

• Before: llama-cli collected reasoning from the stream for display only — only assistant content was stored in messages
• After: reasoning_content is stored alongside assistant content in chat history
• Result: --reasoning-preserve can now re-inject prior thoughts on later turns, enabling proper multi-turn reasoning chains

Prebuilt binaries cover 20+ platform configurations:

If you’re running multi-turn conversations with reasoning models, --reasoning-preserve now works as intended. Prebuilt binaries and source on the b10219 release page.