fix(test): enforce eager mode in Eagle3 quantized E2E test (#715)

## Summary

- Add `enforce_eager=True` to `test_vllm_engine_eagle3` to work around a
vLLM bug where CUDA graph capture crashes with an illegal memory access
on quantized speculator models
- This also fixes the cascading DFlash OOM failure (the Eagle3 crash
leaves a ghost GPU process that eats into the 0.9
`gpu_memory_utilization` budget)

## Root Cause

From the [nightly CI
run](https://github.com/neuralmagic/llm-compressor-testing/actions/runs/28632521105)
(vLLM 0.24.0, Python 3.11):

**Failure 1** — `test_vllm_engine_eagle3[llama3-converted-quantized]`:
vLLM's `EngineCore` crashes during speculator prefill CUDA graph capture
(`cudagraph_utils.py` → `torch.cuda.graph()` →
`torch.cuda.synchronize()` → `CUDA error: an illegal memory access`).

**Failure 2** — `test_offline_smoke[Qwen/Qwen3-0.6B-sharegpt-dflash]`:
OOM during warmup because the ghost process from the Eagle3 crash still
holds 1.83 GiB on GPU, leaving only 4.54 GiB free vs the 4.64 GiB
needed.

## Test plan

- [ ] Nightly CI passes both
`test_vllm_engine_eagle3[llama3-converted-quantized]` and
`test_offline_smoke[Qwen/Qwen3-0.6B-sharegpt-dflash]`
- [ ] File upstream vLLM issue for speculator CUDA graph capture on
quantized models

Fixes: INFERENG-8164

🤖 Generated with [Claude Code](https://claude.com/claude-code)

Signed-off-by: Orestis Zambounis <orestis.zambounis@gmail.com>
Co-authored-by: Claude Opus 4.6 <noreply@anthropic.com>

Author: orestis-z

tests/e2e/regression/test_eagle3_conversion_acceptance.py

@@ -101,5 +101,6 @@ def test_vllm_engine_eagle3(
             tmp_path=tmp_path,
             prompts=prompts,
             acceptance_thresholds=acceptance_thresholds,
+            enforce_eager=True,
             ignore_eos=True,
         )