[XPU] Harden fused GDN recurrent state updates

[gc-fu]

Summary

• validate fused GDN tensor devices, dtypes, shapes, strides, dimensions, and state indices before launch
• replace bounds-unsafe aligned 16-element scalar loads with true scalar transactions
• remove the inline conv_state shift that can race across HV work-groups
• always run the state shift after the main kernel with an explicit SYCL event dependency, including graph replay
• annotate all mutated state/output tensors in the Torch operator schemas
• update the end-to-end sequential GDN test for the current metadata API

Correctness and performance tradeoff

The prior small-grid shortcut could update conv_state while another work-group was still reading the old state. This PR always uses a second state-shift kernel to make ordering correct. Small workloads therefore gain one kernel submission; larger workloads already used a separate shift kernel. The dependency is explicit rather than relying on queue ordering.

Scope

This is a kernel-only hardening stage following #569. The operator names and call signatures are unchanged, so no framework PR is required.

Testing

Not run as part of PR creation.