CUDA: Silence unused-variable warnings in the CUDA backend

[anavp-nvidia]

Overview

Building the CUDA backend currently produces an enormous build log that is largely repeated "unused variable" warnings (nvcc #177-D (declared but never referenced) and nvcc #550-D (set but never used)). This PR silences the warnings by either:

• Narrowing the declarations' scope by moving them inside the if constexpr block (or similar) if the variable is only used there.
• Using [[maybe_unused]] attribute if the variable is used in more than one conditionally-compiled block.

Requirements

• I have read and agree with the contributing guidelines
• AI usage disclosure: Yes, AI tools were only used for code review.

[ggerganov]

Thanks @anavp-nvidia

[ggerganov]

It seems the warnings disappeared in the CUDA 13.3 builds, but they remained in the CUDA 12.4 builds. Any idea why that would be?

• 13.3: https://github.com/ggml-org/llama.cpp/actions/runs/26803101548/job/79023354873#step:7:96
• 12.2: https://github.com/ggml-org/llama.cpp/actions/runs/26803101548/job/79023354765#step:7:293

[JohannesGaessler]

Inconsistent handling of INFINITY, for practical purposes we can just replace it with the negative max. values representable with FP32/FP16 for a block-wide maximum reduction.