WebGPU: Support int32 and uint32 for CumSum (#31702)
### Description
Adds int32 and uint32 data type support to the Cumsum operator in the
WebGPU execution provider. The op is used by
[Phi-4-mini](https://huggingface.co/webnn/Phi-4-mini-instruct-onnx-webnn/blob/main/onnx/model.onnx)
(GQA decomposition).
CumSum's WebGPU kernel only advertised float types (float16/float32), so
int32/uint32 inputs silently fell back to the CPU EP. Widen the T type
constraint to WebGpuSupportedNumberTypes() (adds int32/uint32).
### Motivation and Context
No shader or C++ dispatch change is needed. The kernel is not templated
and the shader is written entirely against the generic value-type alias
(output_value_t): the accumulator is `var sum : output_value_t = 0` and
the reduction is a plain `sum = sum + ...`. WGSL resolves the `0`
literal and `+` for i32/u32 exactly as for the float case (integer add
wraps, matching the CPU CumSum reference), and the loop bounds/indices
are already i32/u32 and independent of the element type. So enabling the
two integer types is purely a matter of relaxing the type-constraint
whitelist.
