DeepSeek V4.1 native MLX runtime

This directory contains code derived from PipeNetwork/deepseek-v41-mlx,
revision 8bdd543c7160800f3451d9595c996129b7abecdf, licensed under the
Apache License 2.0. Rapid-MLX modifications include quantized group-aware
output projection support and qualification tooling.

The DSpark draft algorithm in dspark.py follows the model publisher's
MIT-licensed DeepSeek V4.1 reference inference implementation. The Rapid-MLX
implementation adds a strict data-only sidecar loader, target weight sharing,
memory admission, validation, and benchmark integration.
See LICENSE-DSPARK in this directory.

The hyper-connection Metal kernels in hyper_connections.py include modified
MIT-licensed code from jundot/omlx revision
75466d33a3f9e1bf4b8257738bcadd4658526bcb. See LICENSE-DSPARK in this
directory for the MIT license text.
