[TVM][ARITH] Enhance BoundDeducer

[derisavi]

Consider the following reduced testcase:

import tvm

a = tvm.var('a')
b = tvm.var('b')
b_s = tvm.arith.intset_interval(2, 3)

print(tvm.arith.DeduceBound(a, b < a, {b: b_s}, {}))
print(tvm.arith.DeduceBound(a, a > b, {b: b_s}, {}))

The output is:

interval-set[pos_inf, neg_inf]
interval-set[(b + 1), pos_inf]

The expression in the two cases are mathematically equivalent, i.e., a > b, but because TVM expects the part of the expression that contains the target var (i.e., a) to be on the lhs, it cannot handle the first DeduceBound properly.