Unable to use same variable for reduction twice or more in TC

[prigoyal]

do we not allow people to use same variable twice? also repro is available in #61

FAILS:
def softmax(float(N, D) I) -> (O, tmp) {
        tmp(n) max=! I(n, d)
        O(n, d) = exp(I(n, d) - tmp(n))
        tmp(n) +=! O(n, d)
        O(n, d) = O(n, d) / tmp(n)
}

PASSES:
def softmax(float(N, D) I) -> (O, expsum, maxVal) {
        maxVal(n) max= I(n, d)
        expsum(n) +=! exp(I(n, d) - maxVal(n))
        O(n, d) = exp(I(n, d) - maxVal(n)) / expsum(n)
 }

[prigoyal]

cc @jekbradbury

[prigoyal]

the reason is because we use Halide IR and it makes assumption that the tensor dependencies are strictly DAG. Hence it doesn't work in case of softmax variants that failed.