Construct a Lucid tensor from any object exposing the DLPack protocol (__dlpack__) or from a raw PyCapsule.

Two dialects, picked by what the producer says it is:

• Metal (__dlpack_device__() == (8, 0)) — adopted natively. The producer's MTLBuffer becomes this tensor's storage with no copy at all, which is what makes an mlx.core.array and a Lucid Metal tensor two views of one allocation. NumPy cannot read a capsule of this device type, so this path bypasses it entirely.
• Everything else — through NumPy, which shares host memory where it can. The result lives on the CPU regardless of the producer's device; call .to("metal") after importing.

Requires numpy for the host dialect — install via pip install lucid[numpy] if missing. The Metal dialect needs nothing.