Mandate output_channels % groups == 0 validation for conv2d #925

Background
Currently, the WebNN specification for conv2d ensures that the input channels are evenly divisible by the groups attribute (e.g., verifying that the filter's input channel dimension matches input_channels / groups). However, it lacks a symmetric validation step explicitly ensuring that the output channels are also perfectly divisible by groups.

The Issue
Without this explicit validation mandated by the specification, backend engines may encounter mathematically invalid graphs that lead to undefined behavior. If a developer provides an output_channels value that is not a multiple of groups, it frequently triggers integer truncation bugs in the backend's channel-to-group assignment logic.

Proposed Specification Change
To protect all underlying machine learning frameworks from this class of malformed tensor, we propose adding an explicit validation step to the WebNN spec for conv2d.

The specification must mandate that the total number of output channels is a multiple of the groups attribute.

Motivation
Enforcing this strictly at the WebNN API level guarantees that all backends—whether TFLite, DirectML, or CoreML—are shielded from processing mathematically impossible grouped convolutions. This eliminates an entire class of potential buffer overflows before the graph is ever compiled.

Future Consideration
While convTranspose2d does not currently support groups > 1, this exact same validation rule will be critical to implement for transposed convolutions if group support is added in the future, as an integer division flaw there would result in an Out-Of-Bounds Write.