Fix integer overflow in tensor dimensions and PPM parser

[sharadboni]

Summary

• Tensor dimension overflow (util/basics.h): Extents2D::Area() computes rows * cols without checking for size_t overflow. When tensor dimensions are deserialized from a model file via MatPtr::VisitFields(), attacker-controlled rows and cols values can silently overflow, causing undersized allocations and subsequent heap buffer overflows. Added an overflow check that aborts if cols > SIZE_MAX / rows.

• PPM image size overflow (paligemma/image.cc): ReadPPM() computes data_size = width * height * 3 without overflow protection. A crafted PPM file with large width/height values can overflow this computation, leading to an undersized data_ vector and out-of-bounds writes during pixel parsing. Added a two-step overflow check before the multiplication.

• PPM integer parsing overflow (paligemma/image.cc): ParseUnsigned() accumulates decimal digits into a size_t without detecting overflow, allowing silently wrapped values to bypass subsequent size checks. Added overflow detection that returns nullptr (parse failure) when the accumulated value would exceed SIZE_MAX.

Test plan

• Verify existing tests still pass (bazel test //...)
• Confirm PPM files with normal dimensions load correctly
• Confirm malformed PPM with overflow-inducing dimensions is rejected gracefully
• Confirm model files with overflow-inducing tensor dimensions trigger abort

[sharadboni]

@jan-wassenberg Could you review this security fix? It adds overflow checks for tensor dimensions in Extents2D::Area() and fixes integer overflow in the PPM image parser that can be triggered via crafted model/image files.

[jan-wassenberg]

Thanks, this seems reasonable to rule out. We're only able to land patches targeting the dev branch, would you mind updating the PR?
Also one code suggestion.

[jan-wassenberg]

Thanks for updating! (FYI HWY_ABORT does not return, but we can land as-is)