[bug] resume_from is ignored in sdxl_train_textual_inversion.py

I'm encountering a critical issue with the `resume_from` parameter when training a *textual inversion* embedding on **Stable Diffusion XL** using `sdxl_train_textual_inversion.py`.

Despite specifying a valid `.safetensors` file in the `resume_from` field (with proper token name and path), the training **starts from scratch**: it initializes the embedding as if no prior progress existed.

**Context:**
- I’m using the bmaltais-style config (`.toml`), launched with:
`accelerate launch ... sdxl_train_textual_inversion.py --config_file=/path/to/file.toml`
- The file specified in `resume_from` exists, is readable, and contains a learned token trained for ~2000 steps
- Training resumes at **step 0**, with no error, and the loss is identical to the initial run (~0.105)
- I also tried `pretrained_embedding_path` — same behavior
- I cannot use `resume_from` to continue a long training (10k+ steps), making any interruption fatal

**Expected behavior:**
Training resumes correctly from the `.safetensors` file, restoring the learned vectors and optimizer state.

**Actual behavior:**
The script silently ignores the checkpoint and resets all training state (including token weights and optimizer).

**To reproduce:**
1. Train a textual inversion embedding with SDXL
2. Save a `.safetensors` at step N
3. Stop training
4. Restart with `resume_from = "path/to/embedding-N.safetensors"`
5. Observe that training restarts from step 0, reinitializes embedding, and behaves like a fresh run

**System:**
- GPU: AMD RX 7900 XTX (ROCm 6.3.2)
- Tested with: latest main branch of `sd-scripts`
- Config: bmaltais `kohya_ss` style config + `sdxl_train_textual_inversion.py`

Let me know if you need logs, test cases, or can guide me to patch this. Thanks!

Start in sdxl_train_textual_inversion.py and trace how resume_from and pretrained_embedding_path are read when the bmaltais-style TOML config is passed. Reproduce the documented stop-and-restart sequence with a .safetensors embedding, then verify that training does not return to step 0 or reinitialize the learned token and training state.

Written by the indexing model from the issue text.