Simplified Thunderkittens Port

[Willy-Chan]

Thunderkittens backend support, but with a simplified load_inline() implementation.

This turns out to be much simpler than using separate files, but has its own tradeoffs and implications.

Also, users don't have to put the TK repo in the root directory, it's automatically cloned to the Modal.

[willhu-jpg]

We should add instructions for installing the TK dir in local environments to the README. Think there's also a mismatch with the CUDA version in the current TK repo too. It's hard coded to 12.6 if we rely on TK instructions to call 'source env.src'

[Willy-Chan]

We should add instructions for installing the TK dir in local environments to the README. Think there's also a mismatch with the CUDA version in the current TK repo too. It's hard coded to 12.6 if we rely on TK instructions to call 'source env.src'

@willhu-jpg Added instructions to the README, lmk what you think. Also, could you clarify the CUDA version point? Like the tk-v2 branch uses a different version than kernelbench right now?

[willhu-jpg]

README changes look good. For the CUDA version, checkout the env.src in the TK repo. Typical setup is to call "source env.src" to set all the environment variables, but they have it hardcoded to 12.6.

[willhu-jpg]

changes look good otherwise! let's merge it :)

[willhu-jpg]

nit typo

[Willy-Chan]

Sounds good, fixed the typo. @simonguozirui TK works on CUDA versions 12.6 - 12.9, rn in modal scripts the system uses CUDA 12.8 by default which should be fine. Locally users will set CUDA_HOME, default is 12.6 for TK's env.src. Lmk if this arrangement sounds fine.