Fix issue with corrupted output on long generation for GPT

[andrewchernyh]

Fix issue described in #2300

Only updating MAX_OUT_TOKES doesn't helps and main issue is in temp_buf that was allocated just after output tensor

Then in attention_unfused
T* workspace = (T*)output + bsz * seq_len * heads * k; where output is temp_buf from ds_softmax_context
In the result temp_buf overlaps query_cont allocated in ds_softmax_context

I've put temp_buf just after kv_cache and it helps

Minimal steps to reproduce:

python3 benchmarks/inference/gpt-bench.py -m EleutherAI/gpt-neo-125M --kernel-inject --deepspeed --dtype=fp32 --max-tokens=1020 --trials 1

[ghost]

All CLA requirements met.

[andrewchernyh]

Not actual more, see #2300 (comment)