Improve shape inference with `DecodeAndCropJpeg` (#13750)

* Improve shape inference with `DecodeAndCropJpeg`

While working on improving shape inference for several
other ops in 13561 and 13193, I noticed that `DecodeAndCropJpeg`
does not inference shape even though crop size might have already
be provided. In that case the shape will be `[h, w, channel]`
and `h`, `w` is part of the `crop_window`.

This fix updates the shape function in `DecodeAndCropJpeg`
for improving shape inference.

Signed-off-by: Yong Tang <yong.tang.github@outlook.com>

* Add test cases to cover shape inference for `DecodeAndCropJpeg`

Signed-off-by: Yong Tang <yong.tang.github@outlook.com>

* Address failed unit tests

Signed-off-by: Yong Tang <yong.tang.github@outlook.com>