Add Qwen3

[kanpuriyanawab]

This PR adds Qwen3 Backbone to library

What's pending?
Backbone Outputs are matching with atol 1e-2

very small percentage of output logits fall into 1e-2 and 1e-3 thereby causing assert to fail

^ this is with 0.6 B checkpoints, with 4B checkpoint, it nicely passes with atol 1e-3

[kanpuriyanawab]

I checked multiple times and there is no implementational diff bw keras and hf one except us using einsumdense layers..

Query dense layer causes very small diversion but it is amplified by norm layer (and this norm layer fyi, has been used in all other llms -- llama, qwen2.5, and qwen-moe exact copy paste).. i think culprits are weights

@divyashreepathihalli