feat: add support for DeepEP in Qwen3 parallelization logic

[jordisassoon]

This PR allows to pass expert_parallel_comm_backend = "deepep" and launch experiments with DeepEP on Qwen3 model architectures. Default is expert_parallel_comm_backend = "standard".

Changes are duplicated logic from the deepseek_v3 args/model/infra implementation.

I ran some benchmarks to verify speed-up with expert_parallel_comm_backend = "deepep" compared to "standard", avergaging TPS after warmup:

Benchmark info

8xB200 GPUs

[parallelism]
data_parallel_replicate_degree = 1
data_parallel_shard_degree = 8
tensor_parallel_degree = 1
pipeline_parallel_degree = 1
context_parallel_degree = 1
expert_parallel_degree = 2

[shuhuayu]

Thanks for the contribution. LGTM!