stream fix: read1() not read(65536) — flush SSE as it arrives

_stream_response did resp.read(65536), which blocks until 64KB or EOF. A whole chat
response is only a few KB, so it never filled the buffer and only released at EOF —
reasoning + content arrived in one burst at the end instead of streaming. Upstream
(bosch sglang) streams fine (reasoning 0.2->6.5s); the proxy was the buffer. read1()
returns on first available data. Verified reasoning now streams through :8082. Also
fixes streaming for OWUI.

Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>

llm-proxy.py

@@ -659,7 +659,11 @@ class Handler(http.server.BaseHTTPRequestHandler):
         bytes_out = 0
         try:
             while True:
-                chunk = resp.read(65536)
+                # read1(): return as soon as ANY data is available (one underlying read),
+                # instead of read()'s "block until 64KB or EOF" -- otherwise a small
+                # streaming response (a few KB of SSE) only flushes at EOF, so reasoning
+                # + content land in one burst at the end instead of streaming through.
+                chunk = resp.read1(65536)
                 if not chunk:
                     break
                 self.wfile.write(chunk)