deepseek_v32 chat template: accept wrapper-injected kwargs

[lixiangnlp]

Summary

TokenizerWrapper.apply_chat_template (in mlx_lm/tokenizer_utils.py) sets
kwargs[\"enable_thinking\"] = self.has_thinking on every call and may also
forward tokenize= and similar kwargs through to a registered chat-template
module. mlx_lm/chat_templates/deepseek_v32.py:apply_chat_template currently
hands **kwargs straight to encode_messages, which only accepts a fixed
set of named arguments — so the very first call from the wrapper raises:

```
TypeError: encode_messages() got an unexpected keyword argument 'enable_thinking'
```

This means anyone setting `chat_template_type: "deepseek_v32"` in
`tokenizer_config.json` today is dead in the water before generating a token.

Changes

• Accept `enable_thinking` explicitly and translate it to `thinking_mode`,
  with `thinking_mode` winning if both are set.
• Filter `**kwargs` down to the set `encode_messages` accepts
  (`thinking_mode`, `context`, `drop_thinking`, `add_default_bos_token`,
  `tools`).
• Strip the trailing `<｜Assistant｜>` marker when
  `add_generation_prompt=False` and `enable_thinking=False`, mirroring the
  existing `<｜Assistant｜>` strip for the thinking case.

Tests

New `tests/test_chat_templates.py` covering:

• `enable_thinking=True` → ends with `<｜Assistant｜>`
• `enable_thinking=False` → ends with `<｜Assistant｜>`
• explicit `thinking_mode` overrides `enable_thinking`
• unknown kwargs (`tokenize`, `return_tensors`, …) are silently filtered
• `add_generation_prompt=False` strips the assistant marker in both modes

```
$ python -m unittest tests.test_chat_templates
.....

Ran 5 tests in 0.000s
OK
```

Notes

This is independent of the V4 forward-pass work happening on this branch —
it's a pre-existing latent bug in the `chat_template_type` integration that
becomes load-bearing once anything actually wires `deepseek_v32` as a
fallback (see follow-up PR for fallback-by-model-type).