Audit the use of constant(value, element_type) for quantized types

[sdasgup3]

In the spec, there are several places like this, where we use constant(some_constant_value, element_type(result)) to produce a constant of type element_type(result). However, this should not be confused with the StableHLO constant op which only produces constant with tensor types.

IMO, to avoid confusion, we should introduce a separate meta function, under the spec section "function on values", as follows.

* `constant(x: Value, type: TensorElementType | QuantizedTensorElementType)` is
defined on constant inputs and returns a constant of type `TensorElementType` or
`QuantizedTensorElementType`.

[burmako]

"However, this should not be confused with the StableHLO constant op which only produces constant with tensor types". I am a bit confused because #1647 (the discussion in which seems to have prompted creating this ticket) changes ConstantOp to support constants with quantized tensor type as well?

[sdasgup3]

Per the analysis in here, the following ops ops has usages of constant meta-functions which are yet to be reviewed.

• batch_norm_grad and batch_norm_training which are deferred until Finalize the spec for quantized AllReduceOp, BatchNormTrainingOp, BatchNormGradOp and ReduceScatterOp #1666, and
• epsilon calculation in all the batch_norm_inference, batch_norm_training, and batch_norm_inference ops (to be addressed as part of Consider introducing QuantizedTensorType based on integer multiplier/shift #1404).