Add native multi-GPU device_map support to TensorDeserializer

[abatilo]

Add native multi-GPU device_map support to TensorDeserializer

feat(serialization): add device_map parameter for multi-GPU tensor loading

Add native multi-GPU support to TensorDeserializer via a new device_map
keyword argument. Supports explicit per-tensor placement via a Mapping
and automatic greedy largest-first balancing via a Sequence of devices.

• Greedy balancer uses deserialized_length (bytes) with a min-heap
• Per-device CUDA streams cached in copy threads
• CPU tensors in mixed maps get dedicated buffers to avoid corruption
• Underspecified CUDA devices resolved up front
• read_numpy_arrays guarded against non-CPU device_map targets
• Default path (device_map=None) avoids per-tensor overhead in hot loop
• 19 new unit tests covering placement, balancing, fallback, edge cases

[abatilo]

This change is part of the following stack:

• Add native multi-GPU device_map support to TensorDeserializer #201 ◀

_{Change managed by git-spice.}