Add xpu case

[wendyliu235]

Summary

This PR ports the recent XPU-related perf-eval changes from perf-eval into wenjun_perf-eval.

The update includes:

• adding XPU workload support for gpt_oss_120b
• extending server startup logic to support the xpu_docker runtime
• updating GPU profiles with the XPU profile and matching cache path changes
• propagating agent tags in Buildkite pipeline generation

Changes

.buildkite/generate_pipeline.py

• pass through agent_tags from GPU profiles into Buildkite step agents
• align k8s plugin handling with the source repo changes

lib/gpu_profiles.yaml

• add the XPU profile
• configure XPU queue, agent tags, image repo, runtime, and HF_HOME
• align MI300X and MI355X hf_home values with the source repo

lib/server.sh

• add xpu_docker runtime handling
• support XPU container startup with /dev/dri, host networking, and optional ZE_AFFINITY_MASK
• pass through HF_TOKEN when present
• update health check to bypass proxy for localhost

workloads/gpt_oss_120b_xpu.yaml

• add the XPU workload definition for openai/gpt-oss-120b
• use TP=4 and the current benchmark concurrency settings from the source repo

Motivation

These changes bring wenjun_perf-eval in sync with the working XPU-related updates already made in perf-eval, so the same workload and runtime flow can be used in this repo.

Validation

• verified that the edited files have no editor-reported errors