Update SD1.5 TI embedding output format

[RyanJDick]

This PR addresses #100 (based on the Discord discussion here: https://discord.com/channels/1020123559063990373/1130288610520875049/1227012392416841789 )

Prior to this change, invoke-training produces TI embeddings for SD1.5 in a format based on diffusers with the main key being the placeholder token:

>>> sd = load_file("bruce_the_gnome.safetensors")
>>> sd.keys()
dict_keys(['bruce_the_gnome'])

This format is currently failing to import into InvokeAI via the Model Manager. I suspect that this is caused by a change in behaviour during the recent MM refactor, but I haven't gone to the effort to try and confirm this.

The old diffusers format is not very popular, so instead I have switched invoke-training to use a different format, which uses "emb_params" as the main state_dict key.

Note that only SD1.5 is affected. The format used for SDXL TI embeddings is unchanged.

[RyanJDick]

Test failures are unrelated to this change and will be addressed in another PR.