[yulangz]

Tensor packing currently checks the byte limit after inserting a tensor and uses torch.cat(...).clone(). This can produce oversized groups from individually small tensors and unnecessarily allocate/copy another whole packed group.

Split a nonempty group before an insertion would exceed max_tensor_size, keeping individually oversized tensors isolated. Remove the post-cat clone: concatenation already owns storage independently of the inputs, including single-input groups. Reconstruction metadata and dtype grouping remain unchanged.

Fixes #128.

Validation:

• Tensor utilities and CUDA IPC tests: 31 passed, including exact bounds, oversized-tensor isolation, noncontiguous/single-input storage ownership after input release, and output-only CUDA allocation peak.
• bash ci/test_cpu.sh: 279 passed, 1 skipped in the local environment. Added the packing and CUDA IPC test files to this suite; the device-memory test skips without CUDA.
• Ruff 0.15.20 formatting and lint checks across the repository passed.

The issue's 256 MiB CUDA reproduction measured a 512 MiB packing peak above live inputs with the old cat().clone(), versus 256 MiB with cat() alone. This change targets allocation/copy overhead and the group-size contract; it does not change the CUDA IPC acknowledgement/lifetime protocol.

[Copilot]

Copilot was unable to review this pull request because the user who requested the review has reached their quota limit.