ArraySubbyte::back() returns the wrong element when N does not fill its last storage unit

Description

ArraySubbyte::back() (include/cutlass/array_subbyte.h ~lines 451-458) returns the last storage slot's highest sub-element rather than element N-1:

reference back() {
    return reference(storage + kStorageElements - 1, kElementsPerStoredItem - 1);
}

For any N that does not fill its final storage unit exactly, that slot position is padding, so back() returns an unrelated (often uninitialized) nibble/bit while a[N-1] holds the real last element. Verified by execution:

template <int N> void probe() {
    cutlass::Array<cutlass::int4b_t, N> a;
    for (int i = 0; i < N; ++i) a[i] = cutlass::int4b_t((i % 14) - 6);
    printf("N=%d a[%d]=%d back()=%d\n", N, N - 1, int(a[N-1]), int(a.back()));
}
// N=5:  a[4]=-2  back()=0    WRONG
// N=6:  a[5]=-1  back()=-1   ok      (6 % 4 != 1 case that happens to align)
// N=9:  a[8]=2   back()=-1   WRONG
// N=3:  a[2]=-4  back()=-1   WRONG

front() is correct (at(0)), so the asymmetry is easy to miss.

Same file, two more latent breakages in code paths that fail to compile when instantiated:

• const_iterator's ++/-- (~lines 315-373) declare iterator& return types but build a const_iterator; comparisons take iterator const&, so iterating a const array cannot compile.
• (related family) PredicateVector::Iterator::operator- has the same construct-ConstIterator-return-Iterator shape, reported separately.

Suggested fix

Implement back() via at(N - 1) like front() does for element 0, and give const_iterator self-typed returns/comparisons so it is usable.

Start in include/cutlass/array_subbyte.h around ArraySubbyte::back(), front(), and the const_iterator definitions near lines 315-373. Inspect how at(N - 1) and const iterator operations are implemented, then compile focused uses covering partially filled storage and iteration over a const array. Done means back() returns element N-1 and const iteration compiles with correct self-typed returns and comparisons.