Prevent incorrect CUDA ScatterElements reductions for unsupported types

[Copilot]

Description

• Preserve size-based dispatch for byte-wise reduction="none".
• Dispatch reductions using the tensor’s concrete element type.
• Return an explicit error for types without matching CUDA reduction implementations.
• Add regression coverage for affected integer and BFloat16 types.

Motivation and Context

CUDA ScatterElements reductions interpreted several registered types as same-sized floating-point or signed types, silently producing incorrect results. This change rejects unsupported reductions instead of performing type-punned arithmetic.

• Fixes CUDA ScatterElements reductions dispatch on element size, giving wrong results for several registered types #32061

[azure-pipelines]

Azure Pipelines:
There may be pipelines that require an authorized user to comment /azp run to run.

[Yonghyun Gwon (Noperi0r)]

This overlaps with #29879, which addresses the same CUDA ScatterElements reduction-dispatch issue and already has maintainer approval with passing CUDA-specific CI checks. #29879 supports the affected registered types rather than rejecting them; I’ve linked the overlap in #32061 for coordination.