KSampler (inspire) incompatibility using Euler A

[EgorkaSanyok]

This following behavior happens only with smZNodes installed, vanilla ComfyUI is not affected:

I've been trying to reproduce an image created with the core KSampler, using a workflow that uses the KSampler (inspire) node (https://github.com/ltdrdata/ComfyUI-Inspire-Pack), but I was unable to do it. Upon investigation, it seems that euler_ancestral, dpm_2_ancestral and potentially all other stochastic samplers are the culprits, since I am able to reproduce an image just fine with non-stochastic samplers like euler.

I'll illustrate the issue in the ComfyUI's default workflow, adding only the KSampler (inspire) nodes. noise_mode is set to CPU to match core KSampler, and the rest of the options are set to be identical. To prevent cache copying between nodes, I've changed the batch_seed_mode of both inspire nodes to incremental and comfy respectively, and it shouldn't have an effect on the output since the latent image batch is set to 1.

At first, we notice that core KSampler's output is different from the inspire one, which illustrates the reproducibility issue. However, we then notice that the second inspire node's output is different from the first inspire node. One would expect the output to be at least very similar, even if the samplers are stochastic. If one clears the node caches and runs this workflow again, we'll see that the inspire nodes will produce different images yet again, while the core KSampler stays constant throughout.

To further illustrate this issue, one can cycle through variation_seeds with a variation_strength of 0, which should not impact the output in any way (as it currently behaves in a vanilla ComfyUI install), and observe that whenever euler_ancestral is selected as a sampler, the output is always very noticeably different.

This is the behavior with a vanilla ComfyUI install:

The difference between the core KSampler output in the vanilla install vs the smZNodes install is addressed in #114