Replay provider reasoning state only within the open turn
Stacked on #1035.
Completed turns replay rebuilt reasoning the provider never issued: Anthropic gets a thinking block with no signature, OpenAI gets function_call item ids it did not mint, and the paused rows of a finished turn keep raw blocks forever.
Reasoning state now lives only on steps[].replay_blocks while the turn is paused and is cleared once it completes, so every completed turn rebuilds through generic mapping. DeepSeek reasoning_content is stored as a typed block, and stored tool calls and results carry only provider-neutral keys.
[!WARNING]
Upgrade note for UPGRADE.md: steps[].tool_calls[] no longer stores reasoning_id, reasoning_summary or reasoning_encrypted_content, and thought_signature only when set. StoredMessage::toolResults() and ConversationMessage::$tool_results return id, name, arguments, result, result_id, denied and failed only. DeepSeek AssistantMessage::$replayBlocks becomes a list of ['type' => 'reasoning', 'reasoning_content' => ...] blocks.