Fix DeepSeek-V4 Flash loading and chat

[DiscoStew6082]

Summary

Fix loading and chatting with mlx-community/deepseek-ai-DeepSeek-V4-Flash-8bit.

• Remap DeepSeek-V4 Flash checkpoint keys for quantized top-level embeddings/head tensors, grouped attn.wo_a tensors, and already-stacked routed expert tensors.
• Load TokenizersBackend tokenizers directly from tokenizer.json to avoid AutoTokenizer forcing DeepSeek-V4 config validation.
• Use the existing DeepSeek V3.2 chat template as a DeepSeek-V4 fallback when the tokenizer has no chat template, translating enable_thinking to the template's thinking_mode.

Why

The local DeepSeek-V4 Flash MLX checkpoint failed strict loading with:

ValueError: Received 1423 parameters not in model

After fixing weight remapping, mlx_lm.chat still failed because this model's tokenizer config uses TokenizersBackend and does not include a chat_template.

Verification

Tested locally with:

python -m mlx_lm generate \
  --model /Volumes/Envoy/models/mlx-community/deepseek-ai-DeepSeek-V4-Flash-8bit \
  --prompt "Hello" \
  --max-tokens 1 \
  --temp 0

printf 'Hello, can you talk like a pirate?\nq\n' | uv run --active mlx_lm.chat \
  --model /Volumes/Envoy/models/mlx-community/deepseek-ai-DeepSeek-V4-Flash-8bit \
  --max-tokens 1

Also ran:

python -m compileall mlx_lm/models/deepseek_v4.py mlx_lm/tokenizer_utils.py
git diff --check