litert/xnnpack: reject negative begin-tensor shape dimension in SliceOperation::ToXnnpack()

[destro4evr-rgb]

Problem

PR #11056 added a guard for the 0D scalar case (begin_info.shape.empty()). A 1D begin tensor whose sole dimension is negative (e.g. -1, valid in the flatbuffer format as a dynamic/unknown size marker) bypasses that guard:

if (begin_info.shape.empty()) { ... }   // shape = [-1] → NOT empty, bypassed
size_t num_dims = begin_info.shape[0];  // -1 (int32_t) → SIZE_MAX (size_t)

std::vector<size_t> offsets(num_dims); // requests SIZE_MAX elements → abort
std::vector<size_t> sizes(num_dims);   // same

shape[0] = -1 wraps to SIZE_MAX on the implicit int32_t → size_t conversion. Both vector constructors then request SIZE_MAX elements, exceeding max_size() and causing an unconditional abort at model-load time - before any inference occurs.

Fix

Extend the existing guard to also reject shape[0] <= 0:

if (begin_info.shape.empty() || begin_info.shape[0] <= 0) {
  return absl::InvalidArgumentError(...);
}