[triton_kernels][swiglu] Add standard implementation of swiglu

[matkle]

The current implementation of swiglu computes a variant of the standard SwiGLU activation by adding a bias term (i.e., returning out_gelu * (a_linear + 1) instead of out_gelu * a_linear), and optionally applies input clipping and sigmoid input scaling.

This PR makes the bias addition optional (defaulting to True for backward compatibility) and introduces standard_swiglu, which computes the standard SwiGLU activation without bias addition, input clipping, or sigmoid input scaling.