pub fn freqs_cis(dim: usize, seqlen: usize, base: f32) -> Vec<(f32, f32)>
precompute_freqs_cis(dim, seqlen, original_seq_len=0, base, ...), the no-YaRN form the ratio-0 layers use: [seqlen][dim/2] of (cos, sin).
precompute_freqs_cis(dim, seqlen, original_seq_len=0, base, ...)
[seqlen][dim/2]