Use checked arithmetic for GATHER shape products - #10846

Summary

reference_ops::Gather views the input as [batch, outer, axis, inner] and the positions as [batch, coord]. It computes its batch, outer, inner and flat sizes, and its copy offsets, with int. The GATHER kernel did not check before Eval that any of these products fit. A tensor with a zero dimension has no elements, so BytesRequired accepts it and nothing is allocated. Its other dimensions can still multiply past INT_MAX before the reference code reaches the zero. For example, an input of shape {0, 65536, 32768} gathered on axis 0 gives an inner size of 2^31.

GATHER is on the in-scope operator list in docs/instructions/security.md. This PR follows that document's guidance: use checked arithmetic for every product that feeds a shape, loop bound or offset, and do not let a zero dimension short-circuit that validation.

Changes

• New ValidateGatherFlattenedShapes, called from Prepare after the existing axis and batch_dims checks. It checks the batch, outer and inner sizes of the input, the flat input size, and the size of positions past batch_dims. It uses kernel_utils::CheckedDimensionProduct, the existing helper that ARG_MIN and ARG_MAX use. The helper accumulates in CheckedInt<int> and fails as soon as a partial product overflows, so a later zero dimension does not hide the overflow.
• After ResizeTensor, Prepare checks the flat output size the same way, because reference_ops::Gather also computes it with RuntimeShape::FlatSize(). The constant-input path (noop) and the regular path now share this code. The constant path evaluates the op during Prepare, and it now does so only after the check.

Valid models are not affected. The only newly rejected inputs are zero-element shapes whose other dimensions multiply past INT_MAX. ARG_MIN and ARG_MAX already reject the same kind of shape (see IntermediateInnerProductOverflowRejected in arg_min_max_test.cc).

Tests

New tests in gather_test.cc:

• GatherPrepareTest.Rejects{Inner,Outer,Positions,Output}SizeOverflowWithZeroElements use zero-element shapes in which one of the products that Gather computes is 2^31, one more than INT_MAX. The shapes are chosen so that the product fits a 32-bit size_t, so the overflow is in Gather's own int arithmetic and not in the element count check. The positions and output cases use shapes where only that one product overflows. Each test expects AllocateTensors() to fail.
• GatherPrepareTest.ZeroElementsWithLargeDimensionStillWork checks that {0, INT_MAX} is still accepted and invoked, and that the output shape is unchanged.

PrepareOnlyGatherOpModel now takes optional input and positions shapes. Existing call sites are unchanged.

Testing

CMake build with -DTFLITE_KERNEL_TEST=ON on x86_64 Linux (GCC 13.3, Release):

• gather_test: 70 of 70 tests pass.
• With gather.cc restored to the base version and the new tests kept, the 4 rejection tests fail because AllocateTensors() returns kTfLiteOk. The other 66 tests pass, including ZeroElementsWithLargeDimensionStillWork.