Proposal

Create a new class that hacks Triton autotune to enable reading kernel autotune configuration files from local or remote disks.

Rationale

This can alleviate:

1. potential bwd precision issues on H20
2. compilation failures with certain configurations in some Triton versions
3. cross-task precision alignment
4. faster kernel launch speed

At present, we do not plan to consider automated configuration generation or cross-shape speed optimization; in other words, users will need to manually tune and generate configurations, and fla will provide a script to achieve similar tuning.