ggml_new_tensor_impl: not enough space in the context's memory pool

[daulet]

Looks like some kind of overflow issue similar to this.

ggml_new_tensor_impl: not enough space in the context's memory pool (needed 268443840, available 268435456)

A concrete repro: #321

One has to push the size of a net to reproduce, e.g. in example above it's GPT-2 like net with embedding size 6144 and num of layers 58.

[daulet]

Never mind, turns out I was exceeding the buffer during memory per token estimation, it had hardcoded 256MB initial buffer which I was exceeding.