Bugfix: Add missing data type validations for QDQ ops

[huningxin]

Validate zeroPoint's data type is the same as input's for dequantizeLinear.
Validate scale's data type is the same as input's for quantizeLinear.

Fix: #921

---

Preview | Diff

[fdwr]

Was it not already covered by this step below? Anyway, tis fine to add explicitly to the steps.

If zeroPoint’s dataType is not one of its allowed data types (according to this table), then throw a TypeError.

[fdwr]

👍

[huningxin]

@fdwr

Was it not already covered by this step below?

I suppose it was not covered.

If zeroPoint’s dataType is not one of its allowed data types (according to this table), then throw a TypeError.

For example, zeroPoint data type is int32 which is allowed, while input data type is int8 which is different from zeroPoint's.

[fdwr]

But if dequantizeLinear's input is int32, then according to the table, zero point must be int32 too ("same as input"). So the extra check is redundant If zeroPoint's MLOperand/dataType is not equal to input's MLOperand/dataType, then exception/throw a TypeError. Again though, I don't mind it being there for clarity - merge at will.

[huningxin]

@reillyeon , please take a look, this PR aligns spec steps with Chromium implementation, see details in #921.

[huningxin]

@fdwr @reillyeon , thanks for the review and approvals. I am going to merge it.