Sd-mecha

- fix the handling of negative frequency buckets in `create_filter`
- remove the training loop in `create_filter`, which turned out to be causing issues. the naive cut approach appears to give less burnt out results
- rename parameter `mean` to `alpha` in `crossover` and `distribution_crossover` methods
- change default value of parameter `overlap` in `dropout` to `1.0`, since it is what people should expect by default in general
- optimize `overlap=1.0` in `dropout` to take linear space and time instead of exponential
- clear garbage collector and cuda memory when done merging. it seems to help with merging consecutively many models with a single script

- Minor fix: close safetensors files as soon as merging is completed.

- Fix slerp definitively by testing for nans at the end instead of trying to predict the failure cases.

- use cuda stream for slight speedup of slow merge methods like rotate (~1m saved on sd1.5)

- align the implementation of `ties_sum` with the ties merging paper by 6DammK9
- fix the `slerp` method, which was not working at all