fix(openai): interleave reasoning blocks with paired function calls in conversation history

[Cbrad24]

Fixes #955.

When using OpenAI provider, with OpenAI reasoning enabled models (e.g. gpt-5-mini) with tool calling and conversational history, and a reasoning model decides to plan and execute a multi-tool calling chain (e.g. ListTenants -> ListTenantUsers(tenantId)), the GPT 5 models expect you to return reasoning items that are immediately followed by a paired function_call, and the assistant message last.

This PR reworks the order of returned messages to match that:

Before (current Prism code):
1. { role: "assistant", content: [{ type: "output_text", text: "I found 5 users in tenant..." }] }
2. { type: "reasoning", id: "rs_...fb60..." }
3. { type: "reasoning", id: "rs_...0054..." }
4. { type: "function_call", name: "ListTenants", id: "fc_...fe05..." }
5. { type: "function_call", name: "ListTenantUsers", id: "fc_...0325..." }

OpenAI sees rs_1 followed by rs_2 — rejects it  with error 400 because rs_1 needs its function_call immediately after.

The fix:
1. { type: "reasoning", id: "rs_...fb60..." }
2. { type: "function_call", name: "ListTenants", id: "fc_...fe05..." }
3. { type: "reasoning", id: "rs_...0054..." }
4. { type: "function_call", name: "ListTenantUsers", id: "fc_...0325..." }
5. { role: "assistant", content: [{ type: "output_text", text: "I found 5 users in tenant..." }] }