[tvm][codegen] Make buffer auto broadcast independent to the order of input args

[zhiics]

Currently auto_broadcast of buffer may cause a problem when it hoists the if guard, as the lifted expression may use the arg before it is defined through let binding.

For example, the following code will cause problems during codege, because we will have something like if (o1/x == 1) ... because o1 and x are bound.

    n0, m0, x = tvm.var('n0'), tvm.var('m0'), tvm.var('x')
    n1, m1 = tvm.var('n1'), tvm.var('m1')
    o0, o1 = tvm.var('o0'), tvm.var('o1')

    A = tvm.placeholder((m0, n0), name='A')
    B = tvm.placeholder((m1, n1), name='B')
    C = tvm.compute((o0, o1/x), lambda i, j: A[i, j] + B[i, j], name='C')

    Ab = tvm.decl_buffer(A.shape, A.dtype, name="Ab", buffer_type="auto_broadcast")
    Bb = tvm.decl_buffer(B.shape, B.dtype, name="Bb", buffer_type="auto_broadcast")
    Cc = tvm.decl_buffer(C.shape, C.dtype, name="Cc", buffer_type="auto_broadcast")
    s = tvm.create_schedule(C.op)

This fix hoist the let binding before the uses of the assert/if statements.

Another problem we might need to solve is the the current input arg can only be a single variable instead of an expression, i.e. A = tvm.placeholder((m/x, n), name='A') would fail, @yzhliu has also observed this. We can probably solve this in a follow-up PR. This PR will feed all these args directly in the build statement so that they can be correctly bound.

CC @yzhliu, @icemelon9, @tqchen

[cchung100m]

LGTM

[icemelon]

Thanks @zhiics