MoE decode is slow on Gaudi 2 at low batch size

[NatTuck]

Several modern-ish MoE models (Qwen 3.5/3.6 MoE, Minimax M2.7) seem to be stuck at around 20 tok/s in single session (BS=1) token generation.

I submitted a draft PR ( #1731 ) inspired by the Minimax M3 speedup ( #1673 ) that significantly improves this for the Qwen models (Qwen 3.6 35B-A3B jumps from ~21 to ~45 tok/s at BS=1). The profiling I've managed to do suggests that there's more room for improvement.