[Parallelism Support Matrix Tests] Replace flaky EP relative comparison with hardcoded absolute baseline

[syhuang22]

Description

The previous approach was unreliable because
any upstream vllm change could shift the baseline, causing spurious
failures unrelated to EP performance.

Now the tests compare EP inference time against hardcoded baselines measured on
TPU v7x-8 (512 prompts, Fused: 3.40s, GMM: 2.07s). Tests fail only if
regression exceeds 15%, making them stable and independent of non-EP code paths.

Tests

Both tests verified on TPU v7x-8:

• test_ep_fused_performance: 3.35s (baseline 3.40s, -1.5%) — PASSED
• test_ep_gmm_performance: 2.21s (baseline 2.07s, +6.6%) — PASSED

python -m pytest tests/e2e/test_expert_parallel.py -v -s

# Checklist

Before submitting this PR, please make sure:
- I have performed a self-review of my code.
- I have necessary comments in my code, particularly in hard-to-understand areas.
- I have made or will make corresponding changes to any relevant documentation.