Tag info saved in model metadata ignores keep_tokens_separator

Tag info saved in model metadata ignores `keep_tokens_separator`.

For example with `keep_tokens_separator = ##` and caption `1boy##standing, red` the metadata will save tags `1boy##standing, red` while training will use `1boy, standing, red`.
Metadata should also save `1boy, standing, red` in tag frequency.

Probably other caption modifiers are also broken.

Start by locating the handling of `keep_tokens_separator` and the code that saves tag frequency in model metadata, then compare it with the caption processing used during training. Reproduce the `##` example and verify that metadata records `1boy, standing, red`, while checking whether other caption modifiers follow the same path.

Written by the indexing model from the issue text.