feat: add bidirectional Ministral3 embeddings

[StartupBros]

Summary

Add a bidirectional Ministral3 encoder adapter for SentenceTransformers embedding checkpoints such as nvidia/Nemotron-3-Embed-1B-BF16.

Implementation

• Reuse mlx-lm’s Ministral3 layers and YaRN implementation.
• Replace causal masking with a full bidirectional additive mask that excludes padding keys.
• Return sequence hidden states plus normalized masked-mean embeddings.
• Preserve checkpoint-native tensor names through explicit model.* sanitation.
• Fail closed for causal architectures or non-full-attention layer layouts.

Verification

• Full mlx-embeddings suite: 56 passed.
• Ruff, Black, and git diff --check: pass.
• Real pinned 1B BF16 checkpoint loads and compiles through the oMLX embedding wrapper.
• Four 2048-dimensional MLX vectors compared with normalized sentence-transformers Torch CPU outputs:
  • minimum per-vector cosine: 0.999756
  • all dimensions: 2048
  • module-declared mean pooling exercised

No model repository code is executed, and no checkpoint files are rewritten.