BatchedModelKit does not respect multiple EOS token IDs from model config

Nobody has claimed this yet.

### Summary

`BatchedModelKit` does not currently forward the model's configured `eos_token_id` values when loading the tokenizer.

This breaks models that define multiple EOS / EOG tokens. In such cases, batched generation may continue past a valid end-of-generation token into the next serialized conversation turn.

I reproduced this with `lmstudio-community/GLM-4.7-Flash-MLX-8bit`.

### Reproduction

Using the LM Studio OpenAI-compatible API:

```bash
curl http://localhost:1234/v1/chat/completions \
-H "Content-Type: application/json" \
-d '{
"model": "zai-org/glm-4.7-flash",
"messages": [
{
"role": "user",
"content": "Answer in one word: what is the capital of France?"
}
],
"temperature": 0.2,
"max_tokens": 300
}'
```

Without a manually configured stop string, the response continues after the assistant answer:

```text
Paris<|user|>...
```

and eventually finishes with:

```text
finish_reason: length
```

Adding:

```json
"stop": ["<|user|>"]
```

makes generation terminate correctly with:

```text
Paris
```

and:

```text
finish_reason: stop
```

Direct generation through `mlx_lm.generate` also terminates correctly, so the issue appears specific to the batched engine path.

### Root cause

The model defines multiple EOS token IDs:

```json
"eos_token_id": [
154820,
154827,
154829
]
```

where one of the additional EOS/EOG tokens corresponds to `<|user|>`.

`BatchedModelKit` loads the tokenizer without forwarding these configured EOS IDs:

```python
self.tokenizer = mlx_lm.tokenizer_utils.load(self._model_path)
```

As a result, the tokenizer can fall back to its default/single EOS token instead of preserving the complete model-defined EOS set.

There is also a related configuration difference: `BatchedModelKit` reads `config.json` directly, while `mlx_lm` uses `load_config()`, which also applies relevant values from `generation_config.json`.

### Proposed fix

Use the same configuration path as `mlx_lm` and explicitly forward the configured EOS token IDs:

```python
config_json = mlx_lm.utils.load_config(model_path)

self.tokenizer = mlx_lm.tokenizer_utils.load(
self._model_path,
eos_token_ids=config_json.get("eos_token_id", None),
)
```

This preserves multiple EOS/EOG tokens and respects `generation_config.json` overrides.

### Verification

With this change applied locally, GLM-4.7-Flash stops correctly at `<|user|>` without any manually configured LM Studio Stop Strings.

I also added regression tests covering:

* multiple `eos_token_id` values from the model configuration;
* `eos_token_id` overrides from `generation_config.json`.

The targeted `BatchedModelKit` scheduler tests pass after the change.

### Fix / reference implementation

I prepared the fix and regression tests in my fork:

https://github.com/iphizic/mlx-engine

I cannot currently open a pull request against this repository, so I am providing the implementation here for reference. I would be happy to adjust the patch if a maintainer prefers a different approach.

Begin with BatchedModelKit's tokenizer initialization and compare it with mlx_lm.utils.load_config and tokenizer_utils.load. Run the targeted BatchedModelKit scheduler tests, including cases for model-config and generation_config.json EOS IDs. Done means multiple EOS/EOG IDs are preserved and the GLM reproduction stops at the user token without a manual stop string.

Written by the indexing model from the issue text.