fix: bound tensor header length during deserialization

[dfgvaetyj3456356-hash]

Summary

Bounds each tensor header length against the metadata span that already describes where that tensor's data begins.

Before this change, _TensorHeaderDeserializer.from_io() trusted the 8-byte header length stored in the file and allocated a bytearray(header_len) before checking whether that length fit the metadata entry for the tensor. A malformed .tensors file could therefore make the lazy deserializer allocate based on an attacker-controlled header length, and small oversized headers could also be accepted because missing bytes were effectively padded in the header buffer.

Fix

• Add an optional max_header_len guard to _TensorHeaderDeserializer.from_io().
• In _copy_thread(), derive the legal header span from the metadata entry at the current file offset: entry.data_offset - entry.offset.
• Reject unexpected header offsets and oversized header lengths before allocating the header buffer.

Tests

• python -m py_compile tensorizer\serialization.py tests\test_serialization.py
• python -m pytest tests\test_serialization.py::TestSerialization::test_oversized_tensor_header_is_rejected -q
• python -m pytest tests\test_serialization.py::TestDeserialization::test_lazy_load -q
• python -m pytest tests\test_serialization.py::TestSerialization::test_oversized_tensor_header_is_rejected tests\test_serialization.py::TestDeserialization::test_lazy_load -q
• git diff --check

test_lazy_load emits the existing Hugging Face cache symlink warning on Windows, but passes.