[SM100] Add Kimi-K3 SiTU activation

[Fridge003]

Summary

• add Kimi-K3 SiTU selection to the SM100 FP8/FP4 Mega-MoE L1 epilogue when kActivationClamp == 0.03125f
• use the K3 constants beta = 4.0 and linear_beta = 25.0, bypassing the ordinary SwiGLU hard clamp for the SiTU path
• preserve the existing limited and unlimited SwiGLU behavior for every other clamp value

Why

SGLang's Kimi-K3 integration currently applies this change at container startup through apply_deepgemm_situ_patch.py. Carrying the behavior in DeepGEMM removes the need to rewrite the installed JIT header at runtime. The sentinel clamp value also creates a distinct JIT template instantiation, leaving cached SwiGLU kernels unaffected.

Impact

Callers that pass the Kimi-K3 sentinel get beta * tanh(gate / beta) * sigmoid(gate) multiplied by linear_beta * tanh(up / linear_beta). Existing callers continue to use the original SwiGLU path.

Validation

• compared the modified epilogue block byte-for-byte with the NEW block in the linked SGLang patch
• verified the SiTU sentinel appears exactly once
• git diff --check
• SM100 JIT compilation and GPU numerical tests were not run locally because the current environment has no CUDA runtime or GPU