fix(data): reject raw tensor whose data length != declared shape (#2390)

[czoli1976]

Summary

Fixes #2390. Tensor::from_raw_dt_align allocated ∏(shape) × sizeof(dt) bytes and then copy_from_slice-d the caller-supplied content into it with no length check. When the two lengths differ — e.g. a crafted or malformed .onnx whose initializer raw_data byte count does not match its declared dims — copy_from_slice panics, aborting model load in both debug and release builds. Since tract loads externally-supplied ONNX, this is a reachable DoS.

This validates the content length against the freshly-allocated tensor and returns a TractResult error instead of panicking. The check sits at from_raw_dt_align, the single choke point for all from_raw callers (ONNX, NNEF, …), so every raw-load path is covered.

Reproduction (before this PR)

The 94-byte PoC from #2390 (raw_data of 12 bytes against a shape needing 440):

thread '...' panicked at data/src/tensor.rs:502:31:
copy_from_slice: source slice length (12) does not match destination slice length (440)

After this PR the same file loads to a graceful Err (no panic), verified in release.

Change

let mut tensor = unsafe { Tensor::uninitialized_aligned_dt(dt, shape, align) }?;
let expected = tensor.as_bytes().len();
ensure!(
    content.len() == expected,
    "Raw tensor data length ({}) does not match shape {:?} of {:?} ({} bytes)",
    content.len(), shape, dt, expected
);
tensor.as_bytes_mut().copy_from_slice(content);

Tests

• New regression test from_raw_rejects_length_mismatch in data (short, long, and exact-match cases).
• cargo test -p tract-data --lib (140) and cargo test -p tract-onnx --lib (10) pass; the end-to-end PoC load now returns Err instead of panicking.

🤖 Generated with Claude Code

[czoli1976]

@kali