revert: upload unmodified spacekaren/chatterbox-turbo-webgpu models (our INT64 fix was breaking them)

Fix: detect literal tensor(int64) type strings, not just type constraint params — fixes Unsqueeze/Reshape/Squeeze/Expand/Pad/ReduceSum/ReduceMean/Resize/ScatterND

Schema-driven fix: ConstantOfShape/Tile INT64 constraints + comprehensive type validation

Fix: forward type propagation resolves unknown tensor types; fixes Concat/Reshape cascade errors

Fix structural inputs with unknown types: safety Cast(->INT64) for Concat/Unsqueeze outputs feeding Reshape/Expand/Slice

Fix Range type mismatches: switch to schema-based type detection covering all ONNX operators

Fix all WebGPU type issues: Shape dtype, structural INT64, Cast insertion for type mismatches

fix: restore INT64 for structural inputs (axes/shape/starts) + remove Shape dtype + regenerate value_info

fix: minimal fix - only remove Shape dtype + regenerate value_info via shape inference

fix: use shape inference for correct value_info types (fixes Shape/Gather INT64 mismatches)

fix: preserve INT64 on Shape outputs + remove dtype attr for WebGPU EP

fix: remove Shape dtype attribute + INT64→INT32 for WebGPU EP compatibility

Fix all int64→int32 type annotations for WebGPU compatibility

initial commit