[Relay][Pass] CanonicalizeCast

[vinx13]

In the case that upcast is multi-refed like:

data(i32) - cast(i8) -cast(i32) - elemwise_op
                                - elemwise_op

if we fuse cast(i32) to the previous op, the sub-function after fusion will have 32-bit output, instead of 8-bit output, which has larger memory access overhead.
In such case, it is better to create a copy of cast(i32):

data(i32) - cast(i8) - cast(i32) - elemwise_op
                     - cast(i32) - elemwise_op

Previously we did special handling in CSE to ignore some common expressions in order to make quantization efficient.
This trick is too fragile for example the recent pr #3233 broke it.
This pass makes it easier as we can transform any expressions into such form that has more efficient fusion result.

cc @tqchen @ZihengJiang @eqy
also feel free to give a better name for this pass

[tqchen]

@ZihengJiang @eqy please help to review this PR

[jroesch]

Can we come up with a better name and put in some docs explaining how this pass works in the text format?

Tianqi wrote this example to me yesterday:

def @main(x: int8) {
    %1 = cast(%x, f32)
    %2 = cast(%x, f32)
    %3 = exp(%1)
    %4 = log(%2)
    (%3, 4)
}

then

def @main(x: int8) {
    %1 = fn (%p1: i8) {
        exp(cast(%p1, f32)
    }
    %3 = %1(%x)
    %2 = fn (%p1: i8) {
        log(cast(%p1, f32)
    }
    %4 = %2(%x)
    (%3, 4)
}

[tqchen]

@vinx13 please address review comments. @jroesch can you suggest possible alternatives for naming?

[vinx13]

I have updated this PR. Here are some of the alternative names in my mind: CanonicalizeCast, CanonicalizeFusion.

[tqchen]

CanonicalizeCast seems to be a good name, @jroesch what do you think about it

[jroesch]

I think we should use CannoicalizeCast or DuplicateCast or something since this is a cast focused optimization, its not clear to me whether this is always a performance win or not? is that the case?

[tqchen]

I think it will always be a perf win.

[tqchen]

Thanks @vinx13 @jroesch this PR is now merged