Add drop_keys to drop certain keys from the LoRA network

[rockerBOO]

I was experimenting with dropping certain keys but thought it would be a good experiment to have in the training process. Ideally drop_keys would support regex .*(attentions).*(attn_v).* (one single regex) or a list of keys drop_keys=proj,to_v

I am proposing it as it is but willing to change or extend fully. This implementation does work as one may expect and may be enough to work off from.

network_args=["drop_keys=proj"]

Note: the skipped message should be updated.

because block_lr_weight is 0 or dim (rank) is 0, 140 LoRA modules are skipped / block_lr_weightまた
    lora_te_text_model_encoder_layers_0_self_attn_k_proj
    lora_te_text_model_encoder_layers_0_self_attn_v_proj
    lora_te_text_model_encoder_layers_0_self_attn_q_proj
    lora_te_text_model_encoder_layers_0_self_attn_out_proj
...
    lora_unet_down_blocks_0_attentions_0_proj_in
    lora_unet_down_blocks_0_attentions_0_transformer_blocks_0_ff_net_0_proj
    lora_unet_down_blocks_0_attentions_0_proj_out
    lora_unet_down_blocks_0_attentions_1_proj_in
    lora_unet_down_blocks_0_attentions_1_transformer_blocks_0_ff_net_0_proj
    lora_unet_down_blocks_0_attentions_1_proj_out
    lora_unet_down_blocks_1_attentions_0_proj_in
    lora_unet_down_blocks_1_attentions_0_transformer_blocks_0_ff_net_0_proj
    lora_unet_down_blocks_1_attentions_0_proj_out
    lora_unet_down_blocks_1_attentions_1_proj_in
...

Thank you!