Alternative Skip Layer Guidance (SLG) node implementation

[city96]

This PR adds a version of skip layer guidance (SLG) that only runs two passes, instead of running an extra one (current master). This cuts down on some of the compute overhead[1].

It also matches the logic found in deepbeepmeep/Wan2GP#61 and the results are closer to the outputs of ComfyUI-WanVideoWrapper by @kijai

I was playing around with this for a bit, and was getting better outputs when using it with Wan 2.1 than I was with the current SkipLayerGuidanceDiT node, though it works with other DiT models as well since the code is generic.

Here are some samples for comparison: (imgsli link)

Workflow screenshots

No SLG

Old SLG

New SLG

Kijai SLG

Speed figures for the L40 GPU I was testing on via runpod:

slg mode	speed
None	20/20 [00:50<00:00, 2.53s/it]
Old	20/20 [01:16<00:00, 3.83s/it]
New	20/20 [00:51<00:00, 2.56s/it]

---

[1] I had to add a way to replace calc_cond_batch to do this, because that's the only place where cond/uncond are clearly separated and model_options is exposed.

Other ideas I considered:

• pre_cfg_function - can replace uncond here with a separate pass, but that still has the compute overhead - Could add a force_cfg1_optimization arg to model options that forces skipping uncond, but that could cause issues for other things using pre_cfg_function depending on the order.
• model_function_wrapper applied here - only exposes transformer options, would have to manually add skip block to only the uncond. Separating cond/uncond here is also messy if they're batched.
• Inside the actual model where blocks_replace gets applied - replacement function doesn't have access to transformer_options / same issue with cond/uncond being batched
• Custom guider - too much code duplication, annoying to refactor workflows around custom sampler just to use SLG.

I think it might be better to use a wrapper/hook but I haven't used that system enough to know whether it's possible or not. Tagging @Kosinkadink based on the PR for the original hooks update in case there's a better way, though I think this current version is relatively clean.