Improve L0_client_memory_growth test

[rmccorm4]

1. Refactor memory_leak_test.cc

   • Fix behavior of -R (reuse) flag to actually reuse client/connection. Previously subsequent calls to client::Create(reuse_client. ...) would just create new client objects under the hood.
   • Template helper functions since HTTP and GRPC APIs used are identical in this test

2. Update C++ HTTP client to correctly propogate errors on client-side curl failures. Now the behavior matches other API call failures, rather than reporting an obscure Timer and Json parsing issue. See the before/after examples below:

Before:

root@rmccormick-dt:/workspace# memory_leak_test
Failed to update context stat: Timer not set correctly. Send time from 1682559474332558400 to 0.
error: unable to run model: failed to parse the request JSON buffer: The document is empty. at 0

root@rmccormick-dt:/workspace# simple_http_infer_client
Failed to update context stat: Timer not set correctly. Send time from 1682559478689285651 to 0.
error: unable to run model: failed to parse the request JSON buffer: The document is empty. at 0

root@rmccormick-dt:/workspace# simple_http_health_metadata
error: unable to get server liveness: HTTP client failed: Couldn't connect to server

root@rmccormick-dt:/workspace# simple_http_model_control
error: Failed to get repository index: HTTP client failed: Couldn't connect to server

After:

root@rmccormick-dt:/mnt/triton/jira/4524-client-growth/client/build# ./install/bin/memory_leak_test
error: HTTP client failed [400]: Couldn't connect to server

root@rmccormick-dt:/mnt/triton/jira/4524-client-growth/client/build# ./install/bin/simple_http_infer_client
error: unable to run model: HTTP client failed [400]: Couldn't connect to server

root@rmccormick-dt:/mnt/triton/jira/4524-client-growth/client/build# ./install/bin/simple_http_health_metadata
error: unable to get server liveness: HTTP client failed: Couldn't connect to server

root@rmccormick-dt:/mnt/triton/jira/4524-client-growth/client/build# ./install/bin/simple_http_model_control
error: Failed to get repository index: HTTP client failed: Couldn't connect to server

The root cause of the connection errors above in CI was running out of ephemeral ports on the system. Sockets were being opened too quickly / too many hanging sockets were being left in WAIT state, causing an inability to assign a port to new connections.

For future reference, this can be observed by running watch ss -s in a separate shell while running this script. If you set the -R flag to re-use the same client connection, you will see a consistent flat usage of sockets.

If you omit this flag and create new connections on each request, you will see a consistent rise / high value in sockets stuck in timewait:

$ ss -s

Total: 2031
TCP:   14135 (estab 47, closed 14068, orphaned 0, timewait 14065)

---

Possible Future Improvements:

• Test AsyncInfer
• Separate test to explicitly reuse client connection
• Expose --max-retries rather than hard-coded value
• Only retry on specific error messages like "couldn't connect to server"

---

Corresponding server PR that just adds notes about these findings to help future debuggers: triton-inference-server/server#5710