Use output shape when lowering Unique to krnl

[chentong319]

This issue is discovered by #2799 . It may not be considered a bug in lowering. The issue is triggered by two imperfect components: onnx backend test and buffer deallocation.

1. onnx backend test
   The output shape of Unique OP depends on the content of the input tensor. But the onnx backend test creates the onnx model based on the chosen input. Therefore, the output shape is static. I opened an issue in onnx.

2. Buffer deallocation
   Due to the mismatch of the generated and specified type of the outputs UniqueOp, the existing lowering of Unique created some unrealized_conversion_cast ops. The deallocation pass in the new llvm can not handle the conversion for returned tensors correctly.

Anyway, I try to fix the problem from onnx-mlir with this PR. I use the type of the outputs for allocation if the type is static.

[negiyas]

@chentong319 Your fixes look good to me! Thank you so much for fixing lowering Unique to krnl.

[chentong319]

@chentong319 Your fixes look good to me! Thank you so much for fixing lowering Unique to krnl.

Could you approve the PR?

[hamptonm1]

@chentong319 Your fixes look good to me! Thank you so much for fixing lowering Unique to krnl.

Could you approve the PR?

@chentong319 Since @negiyas said this looks good, I will make sure to merge this in for you :)

[hamptonm1]

LGTM!

[jenkins-droid]

Jenkins Linux s390x Build #14814 [push] Use output shape when lo... started at 14:48

[jenkins-droid]

Jenkins Linux ppc64le Build #13839 [push] Use output shape when lo... started at 15:01

[jenkins-droid]

Jenkins Linux amd64 Build #14809 [push] Use output shape when lo... started at 13:48

[jenkins-droid]

Jenkins Linux amd64 Build #14809 [push] Use output shape when lo... passed after 1 hr 14 min

[jenkins-droid]

Jenkins Linux s390x Build #14814 [push] Use output shape when lo... passed after 1 hr 46 min

[jenkins-droid]

Jenkins Linux ppc64le Build #13839 [push] Use output shape when lo... passed after 2 hr 0 min