Fix packed GatedDeltaNet shape inference (#32708)
### Description
Updates the `GatedDeltaNet` schema to support packed QKV inputs with
omitted key and value inputs.
Shape inference now:
- Accepts rank-2 and rank-3 packed QKV tensors.
- Preserves the packed tensor’s token dimensions.
- Derives output value-head dimensions from the recurrent state.
- Retains existing behavior for separate Q, K, and V inputs.
- Rejects cases where only one of key or value is provided.
Adds focused shape-inference tests for packed and dense token layouts.
### Motivation and Context
Currently, we can emit `GatedDeltaNet` with Q, K, and V concatenated
into one tensor. The existing schema required separate key and value
inputs and only inferred shapes for rank-3 or rank-4 query tensors.
As a result, models using the packed representation failed graph
validation before execution. This change aligns schema validation and
output-shape inference with a packed QKV contract.
---------
Co-authored-by: copilot-swe-agent[bot] <198982749+Copilot@users.noreply.github.com>
