[moe][compile] turn on capture_scalar_outputs and assert fullgraph on grouped experts

[xmfan]

Stacked PRs:

• ->[moe][compile] turn on capture_scalar_outputs and assert fullgraph on grouped experts #1861

---

1. There's a commitment to turn capture_scalar_outputs on by default, and I'm adding sample MoE tests to core. So we can flip this back.
2. using the compile(disable(compile(_forward, fullgraph=True)), fullgraph=False) trick, we can ensure the grouped gemm is always captured.

logs: https://gist.github.com/xmfan/4fd7efc7718795d302304dab807fabd6