Add int32 support for CPU Trilu

[the0cp]

Description

Add INT32 input support for the CPU implementation of ONNX Trilu.

The ONNX Trilu schema allows tensor(int32), and CUDA EP already supports it, but CPU EP currently does not register an INT32 kernel and fails during session creation with NOT_IMPLEMENTED.

• adds int32_t to the CPU ONNX Trilu(14) kernel registration
• dispatches by tensor element type instead of element size
• adds INT32 upper/lower tests
• updates the CPU operator kernel documentation

Motivation and Context

The CPU Trilu implementation previously dispatched based on element size. That works only as long as each supported element size maps to a single supported logical type. Adding INT32 would otherwise make it share the float path because both are 4-byte types.

The old size-based dispatch is ambiguous for types with the same element size, such as float/int32 and double/int64. This PR switches dispatch to the actual ONNX tensor element type so each registered type maps to the matching TriluImpl<T> specialization.

Fixes #21400.

[the0cp]

Hi @skottmckay, sorry for the gentle ping, would you be the right person to review this?