[IntSet] Avoid overly pessimizing while recursively evaluating bounds

[ajtulloch]

Proposed solution for (part of) #3770.

This doesn't solve the whole problem but helps a bit in some cases. I would much prefer a better solution, possibly the one in ajtulloch@ca1088c.

Reproduction:

import tvm

n = tvm.var('n')
X = tvm.placeholder(shape=(n,), name="x")
W = tvm.placeholder(shape=(n + 1,), dtype="int32", name="w")

def f(i):
    start = W[i]
    extent = W[i+1] - W[i]
    rv = tvm.reduce_axis((0, extent))
    return tvm.sum(X[rv + start], axis=rv)
Y = tvm.compute(X.shape, f, name="y")
s = tvm.create_schedule([Y.op])
print(tvm.lower(s, [X, W, Y], simple_mode=True))

Before:

produce y {
  for (i, 0, n) {
    y[i] = 0f
    for (rv, 0, (w[(i + 1)] - w[i])) {
      if ((rv < (w[(i + 1)] - w[i]))) {
        y[i] = (y[i] + x[(rv + w[i])])
      }
    }
  }
}

After:

produce y {
  for (i, 0, n) {
    y[i] = 0.000000f
    for (rv, 0, (w[(i + 1)] - w[i])) {
      y[i] = (y[i] + x[(w[i] + rv)])
    }
  }
}

[tqchen]

Superceded by #4076