feat: Allow patches applied to ModelPatcher to be implemented in a general form as functions.

[ltdrdata]

Allows this style model patching.

        new_model = model.clone()
        seed_offset = 0
        for key in model.model_keys:
            if key.find(target) < 0:
                continue

            def patch_function(p, weight, key):
                tp = weight.dtype
                gen = torch.manual_seed(tilt_seed + seed_offset)
                noise = torch.rand(size=weight.size(), dtype=tp, layout=weight.layout, generator=gen, device='cpu')
                noise = (noise.to(weight.device) - 0.5) * p[0] * 2

                gen = torch.manual_seed(mask_seed + seed_offset)
                mask_noise = torch.rand(size=weight.size(), dtype=tp, layout=weight.layout, generator=gen, device='cpu')
                noise *= ((mask_strength - mask_noise.to(weight.device)) * 1000000).clamp(0, 1)

                return weight + noise

            new_model.add_patches({
                key: ("patch_function", patch_function)
            }, strength_model=1.0, strength_patch=tilt_strength)
            seed_offset += 1

Eliminate the need for code patches like this:
https://github.com/icefairy64/comfyui-model-tilt/blob/main/model_patcher_add_tilt.patch

[shawnington]

Oh man, calculate_weight could use a re-factor, beautiful beautiful spaghetti, but this is a good idea.