fix: buffer non-streaming upstream responses for currency e2e demo

[Spherrrical]

Problem

The e2e-demo-currency CI job fails on advanced/currency_exchange hurl tests even though the demo flow mostly works. Logs show:

• llm_gateway: upstream response parse error: JSON parsing error: EOF while parsing a string on non-streaming OpenAI responses (body truncated mid-JSON)
• prompt_gateway: response body empty, chunk_start: 0, chunk_size: 0 during streaming
• Hurl: transfer closed with outstanding read data remaining on the chat completions request

Root cause

Both WASM filters parsed non-streaming upstream bodies on the first response chunk. When Envoy delivered the JSON across multiple chunks, partial bodies were forwarded/parsed, causing JSON EOF errors and broken client streams.

Fix

• llm_gateway: Accumulate non-streaming response chunks (suppress intermediate chunks) and parse only when the stream ends. Flush any remaining SSE processor/buffer bytes on a final empty EOS chunk for streaming responses.
• prompt_gateway: Same buffering for non-streaming tool-call responses that enrich metadata before returning to the client.
• Test runner: Add advanced/currency_exchange to the hurl retry list (3 attempts), matching preference_based_routing, to tolerate transient external API flakiness.

Testing

• cargo +1.93.0 clippy -p llm_gateway -p prompt_gateway -- -D warnings
• cargo +1.93.0 test -p llm_gateway -p prompt_gateway --lib
• cargo +1.93.0 build --release --target=wasm32-wasip1 -p llm_gateway -p prompt_gateway

Full e2e requires CI (e2e-demo-currency) with API keys and the plano Docker image rebuild.