Adds the `as_dlpack` function pointer to `mts_array_t` in the C API,
implements it for C++ `DataArrayBase`/`SimpleDataArray`/`EmptyDataArray`,
vendors the DLPack header, fixes cmake install and stream parameter type.

Implements `as_dlpack` override for `TorchDataArray` using ATen's
`toDLPack`, with device mapping, versioned wrapper, and stream stubs.

Adds DLPack struct definitions to the Python ctypes layer, implements
`_mts_array_as_dlpack` callback, and adds baseline Python tests.

Implements `as_dlpack` on the Rust `Array` trait using the `dlpk` crate,
switches data storage to `ArcArray<T>` for generic type support, and
improves DLPack version negotiation.

Enables multi-dtype serialization by rewriting the npy reader/writer to
use DLPack tensors. Adds `half` crate dependency, unifies endianness
handling, and adds Python roundtrip tests for all numpy dtypes.

Implement mts_tensormap_load_mmap_partial in Rust and expose through
the C API, C++ wrapper, and Rust bindings. Supports key, sample, and
property filters for selective memory-mapped loading.

Expose load_mmap_partial in metatensor-core Python, metatensor-torch
(including TorchScript op), with tests and documentation.