[Bug] fix steps check about llama2.go

[nikolaydubina]

interesting. @datastream recommend submitting this issue/fix to llama2.c too

[nikolaydubina]

fixed. thanks @datastream 👍🏻

[nikolaydubina]

investigation notes

key_cache len: p->n_layers * p->seq_len * p->dim

if steps is too large
steps == seq_len 

at the end of main loop 
pos == steps - 1
(pos == seq_len - 1)

at the end of loop in transformer 
t == pos (thus, t == seq_len - 1)
float* k = s->key_cache + loff + t * dim + h * head_size;

l == num_layers-1
loff == l * p->seq_len * dim

h == num_heads - 1

then loop up to head size + head_size.

---

thus k  = (num_layers-1) * seq_len * dim + (seq_len-1) * dim + (num_heads-1) * head_size + (head_size - 1)
k = num_layers * seq_len * dim - seq_len * dim + seq_len * dim - dim + num_heads * head_size - head_size + head_size - 1
k = num_layers * seq_len * dim - dim + num_heads * head_size - 1
k = num_layers * seq_len * dim - dim + num_heads * dim / num_heads - 1
k = num_layers * seq_len * dim - 1

max_k index = num_layers * seq_len * dim - 1

dim = 768
num_heads = 12
head_size = dim  / num_heads

---

thus official run.c code is correct and does not breach array size.