feat(kernels): explicit hd512 sm_scale and the final-logit softcap

[FeathBow]

Description

Closes #872

The four hd512 attention wrappers baked in rsqrt(512), and hd512's only prospective consumer cannot use it: Gemma 4 runs unscaled attention (scaling = 1.0 in the HF reference, for both layer kinds). This makes sm_scale an explicit parameter with a finite check, as the hd256 single-prefill wrapper already does. The entries have no current callers, so no call sites move.

Alongside it, the last elementwise piece the 12B forward needs: an in-place final-logit softcap, cap * tanh(x / cap) in f32 with a single rounding back to bf16. Every published Gemma 4 size declares final_logit_softcapping 30.0; nothing in the elementwise family could express it. The wrapper validates the cap positive and finite and the buffer extent under the crate's capacity contract; the launcher computes its grid with an overflow-free ceil-div.

Test Env

Single GPU (sm_89, x86_64), CUDA 12.9, PEGAINFER_REQUIRE_GPU=1.

Verification

• cargo fmt --all -- --check — clean.
• cargo build --release -p pegainfer-kernels --tests — clean.
• cargo clippy --release -p pegainfer-kernels --all-targets -- -D warnings — clean, zero diagnostics.
• cargo test --release -p pegainfer-kernels --test softcap_smoke — running 2 tests, 2 passed: the softcap op against a host f32 reference on constructed cases, and cap validation (rejects zero/negative/non-finite).

Type of Change

• Bug fix (non-breaking change which fixes an issue)
• New feature (non-breaking change which adds functionality)
• Breaking change (fix or feature that would cause existing functionality to not work as expected)
• Documentation update