Validate safetensors data offsets against file boundaries

[matinsaurralde]

Proposed changes

Addresses #3409.

The SafeTensors loader reads data_offsets from JSON metadata without checking whether they fall within the actual file. A crafted .safetensors file can declare offsets that exceed the file size, causing an out-of-bounds memory read when the tensor is evaluated.

This is the file-boundary counterpart to #3364 (which validates internal consistency between data_offsets and shape * dtype). The Rust reference implementation enforces both checks (TensorInvalidInfo and MetadataIncompleteBuffer).

Changes

mlx/io/safetensors.cpp — After parsing the JSON header, compute the available data size from the actual file size and validate each tensor's data_offsets:

• data_offsets has at least 2 elements
• begin <= end
• end does not exceed the data region of the file

mlx/io/load.h — ParallelFileReader::seek() and FileWriter::seek() did not handle std::ios_base::end, silently falling through to SEEK_CUR. Added SEEK_END support and replaced the hardcoded 0 with SEEK_SET.

tests/load_tests.cpp — Added tests for:

• File with data_offsets exceeding file boundary (must throw)
• File with reversed data_offsets (must throw)
• Valid file still loads correctly (regression)

Checklist

• I have read the CONTRIBUTING document
• I have run pre-commit run --all-files to format my code
• I have added tests that prove my fix is effective or that my feature works
• I have updated the necessary documentation (if needed)

[angeloskath]

Thanks for the PR. The original PR did not actually pass the test that it was supposed to be made for, so I would advise compiling and running the tests before PRing.

It is now fixed and the tests unified.