use size_t for reduced input dims in reduce compute kernels

[aizu-m]

Building with -Wshorten-64-to-32:

src/operator-run.c:1435:22: warning: implicit conversion loses integer precision: 'size_t' (aka 'unsigned long') to 'int' [-Wshorten-64-to-32]
 1435 |   int input_shape1 = context->input_shape[1];

xnn_compute_contiguous_reduce and xnn_compute_discontiguous_reduce copy the reduced dimensions out of context->input_shape (a size_t array) into int locals, then use them as loop bounds. A reduced dimension of 2^31 or more truncates to a negative int. In for (size_t i = 0; i < input_shape1; ++i) that negative value converts back to a near-2^64 bound, so the loop walks the input and output pointers well past their buffers. The discontiguous path also hands the truncated values straight to the size_t k2/k3 parameters of discontiguous_reduce2.

Every other extent, stride and index in both functions is already size_t. These four locals were the only ones left as int. Changed them to match.