security: prevent integer overflows in shape multiplication and quantization allocation

[deezsecc]

Summary

This PR addresses the bug reported in #3582

Changes

memory_helpers.cc — Guard against shape dimension overflow

• Added <limits> include.
• Before multiplying element_count by each dimension, validates that the dimension is non-negative and that the multiplication would not exceed std::numeric_limits<int>::max(). Returns kTfLiteError immediately if either check fails.

micro_allocator.cc — Guard against quantization channel count overflow

• Added <limits> and <new> includes.
• Before computing the allocation size for per-channel quantization params (TfLiteIntArray), validates that channels is positive and that sizeof(TfLiteIntArray) + channels * sizeof(int) would not overflow size_t. Returns kTfLiteError with a diagnostic message if the check fails.

Testing

Two new tests added to micro_allocator_test.cc:

• TestTensorShapeIntegerOverflow — constructs a flatbuffer tensor with shape [65536, 65536] (product overflows int) and asserts that InitializeTfLiteTensorFromFlatbuffer returns kTfLiteError.
• TestQuantizationChannelsIntegerOverflow — on 32-bit targets (sizeof(size_t) == 4), crafts a malformed flatbuffer with a fake channel count of 0x3fffffff (triggers TfLiteIntArrayGetSizeInBytes overflow) and asserts that InitializeTfLiteTensorFromFlatbuffer returns kTfLiteError.

BUG=#3582