pub fn softcapped_window_decode_lse_f32(
query: &[f32],
key: &[f32],
value: &[f32],
batch: usize,
key_len: usize,
heads: usize,
kv_heads: usize,
head_dim: usize,
scale: f32,
window_size: usize,
soft_cap: Option<f32>,
) -> (Vec<f32>, Vec<f32>)