feat(streaming): add streaming conversation support

[tharropoulos]

Change Summary

Add streaming conversational search (conversation_stream) for both documents.search and multi_search, in the sync and async clients.

• documents.search_stream() and multi_search.perform_stream() return a context-managed stream. Iterate it for the answer's {conversation_id, message} pieces, then call get_final_response() for the full search response.
• search()/perform() with conversation_stream: True still accept a stream_config (dict or StreamConfigBuilder) with on_chunk, on_complete and on_error, matching typesense-js. They return the final response.
• The events are parsed by a small SSE decoder that follows the WHATWG rules. It splits on CR/LF at the byte level, so CRLF and multi-byte characters split across reads, and U+2028 inside JSON, are handled. httpx's iter_lines gets the last one wrong.
• Streams go through the same node selection and failover as other requests, but only until the response headers arrive. After that, errors are raised from the iterator and never retried, so callers never see an answer piece twice. A stream holds its max_concurrent_requests slot until it is closed.
• New stream_read_timeout_seconds setting (default 60s, the server's LLM timeout). It is used as the read timeout for streams, since the first piece only arrives once the LLM starts answering.
• Adds the conversation, conversation_model_id, conversation_id and conversation_stream search parameters, and the top-level conversation on multi-search responses.

Tested with respx (failover, mid-stream errors, early close, httpx2) and against Typesense 30.2 and 31.0.rc18 with an OpenAI model (-m open_ai).

PR Checklist

• I have read and signed the Contributor License Agreement.