[CI] Add P2pNccl integration test + rename nixl_integration to pd_integration

[eicherseiji]

Summary

• Rename tests/v1/kv_connector/nixl_integration/ to pd_integration/ — the directory now hosts tests for multiple PD connectors, not just Nixl
• Add run_p2p_nccl_accuracy_test.sh for P2pNcclConnector PD accuracy testing on 2 GPUs
• Add CI entry in distributed.yaml (marked optional)

Test plan

• Existing NixlConnector CI tests pass with renamed directory
• P2pNccl accuracy test passes on NVLink GPUs (2 GPU)

[gemini-code-assist]

Code Review

This pull request refactors the integration test directory for PD connectors from nixl_integration to the more general pd_integration and adds a new integration test for the P2pNcclConnector. The changes look good, but I've identified a potential race condition in the new test script run_p2p_nccl_accuracy_test.sh. My feedback focuses on making the server readiness checks more robust to prevent flaky tests by using health check endpoints instead of relying on less reliable methods and fixed sleep durations.

[gemini-code-assist]

The current wait_for_server implementation is not robust. It uses curl to check the /v1/completions endpoint. A GET request to this POST-only endpoint will receive a 405 "Method Not Allowed" response, but curl without the -f flag will still exit with code 0. This can cause a race condition where the script proceeds before the server is fully ready.

A better approach is to use the /health endpoint, which is a GET endpoint, and the curl -f flag to ensure curl fails on HTTP errors. This makes the server readiness check much more reliable and prevents flaky tests.

Suggested change

	wait_for_server() {
	local port=$1
	timeout 600 bash -c "until curl -s localhost:${port}/v1/completions > /dev/null; do sleep 1; done"
	}
	wait_for_server() {
	local port=$1
	local endpoint=${2:-/health}
	timeout 600 bash -c "until curl -fs http://localhost:${port}${endpoint} > /dev/null; do sleep 1; done"
	}

[gemini-code-assist]

Instead of a fixed sleep 5, it's more reliable to actively wait for the proxy server to be ready. The proxy server exposes a /healthcheck endpoint. Using an improved wait_for_server function to poll this endpoint ensures that the script only proceeds once the proxy is fully initialized, avoiding potential race conditions.

Suggested change

	sleep 5
	echo "Waiting for proxy instance on port 8192 to start..."
	wait_for_server 8192 /healthcheck