[BUG/DEPRECATION] Remove fused attention/mlp

[Qubitium]

Reason for PR:

• Broken since at least March 1st with no fix in sight
• No longer have performance value compared to marlin and new marlin kernels (vllm)
• Why fix/maintain something that no one would want to use in production env?

Ref: Issue #655

Changes:

• Remove all fused attention/mlp code

TESTS

• PASS: test_quantization
• PASS: test_serializtion
• PASS: test_triton
• PASS: test_q4

[fxmarty]

Hey @Qubitium sorry for taking time to review. I've had too many other things in my plate, I'll try to review everything by Friday. Do you have an email/slack/twitter I could reach out to you at? I feel like for some non-huge PRs/bugfixes it would be simplier if you had write access to the main branch. What do you think?

[Qubitium]

@fxmarty Sure thing! You can reach me via twitter @ https://twitter.com/qubitium or email at qubitium@lbx.dev. Willing to help out anyway I can with the project. Most importantly, I am willing to put in the grunt work to get all the non-tensor (cuda/triton) math level things working, fixed, and added.