This is to get around evhtp thread-pinning issue in InferenceServerClient
(1) InferenceServerClient health check uses keep-alive connections so stays in thread 0
(2) Model work starts on thread 0 and potentially blocks it for > 60 sec
(3) InferenceServerClient health check tries to reuse the connection on thread 0 and doesn't get a response VS
urllib (or creating a fresh client) would start a new one on thread 1 and get a response instantly

Root cause confirmed by running model loading and health checks in a tight loop, health checks using the following methods
(i) re-using the same health check client (current setup)
(ii) creating a short lived health check client for each request
(iii) urllib

with the default --http-thread-count value (8) only (i) fails, with --http-thread-count 1 all methods fail