Include the expert gradients in the unfused fp16 clip norm

[vineethsaivs]

With MoE and the fp16 unfused optimizer, gradient clipping is decided from a norm built out of the shared parameters alone, so the clip coefficient is too large and every gradient, expert ones included, is under-clipped.

step calls split_params_grads_into_shared_and_expert_params and discards the expert half; step_fused_lamb spends its expert_norm_groups on the overflow check only. The split exists because expert parameters are replicated across the expert-parallel group, so their norm needs a reduction over that group that get_weight_norm cannot do. FP16_Optimizer folds them in with get_norm_with_moe_layers, and this class already carries the norm_type that call wants but never uses it. Fold them in the same way, and pass the engine's has_moe_layers through so it can.

Test: TORCHDYNAMO_DISABLE=1 DS_ACCELERATOR=cpu PYTHONPATH=.:tests python -m pytest tests/unit/runtime/half_precision/ -q. Two new cases fail before, both spellings of the step: four shared gradients of 3.0 and four expert gradients of 4.0 give a reported norm of 6.0 where 10.0 is correct. Two more pin that a run without experts is unchanged. Nine pass and 32 skip after. yapf 0.43.0 and flake8 pass on the changed files. This touches the two lines #8588 also rewrites, so whichever lands first, I will rebase the other. Executed on Apple M2 Pro CPU; no GPU or distributed run. Single rank, so the expert-parallel group is None and the helper runs as plain arithmetic.