implementin an Engine to serve the trained model by inferencing

[Majdoddin]

trafficstars

see #346

[Majdoddin]

for each token, the resulting tensors should be concatenated to the cache tensors, which is a costly O(n) operation. Is it not better to initialize each cache as [max_new_tokens, ...]?

[Majdoddin]

Have just added kv_cache for RMSNorm, to show how it works. @karpathy how do you like it?