[Tracking] Reduce apply() overhead & improve Adapter usability

[YiyanZhai]

Summary

This issue tracks two related improvements in FlashInfer-Bench:

1. Reduce Python-side apply() overhead so it’s negligible compared to kernel runtime.
2. Improve the Adapter API so it’s easier to use.

Motivation

1. apply() overhead

• apply() overhead makes it harder to trust end-to-end latency numbers for very fast solutions.
• The Python orchestration cost around apply() is currently ~2% on Llama 3.1 8B, and can be further reduced.

2. Adapter usability

• Writing a new Adapter currently requires understanding several internal concepts like dispatch workflow.
• We’d like a smoother path for:
  • Adding a new adapter.
  • Configuring existing adapters.

[YiyanZhai]

Apply overhead reduction: #121