Add cuda-oxide single-item DWT97 transcode kernels

[jcwal1516]

Summary

• extend the cuda-oxide transcode PTX project with single-item irreversible 9/7 IDCT, row lift, and column lift kernels
• route only reversible 5/3 and single-item irreversible 9/7 transcode through cuda-oxide when the feature/env gate is enabled
• keep batched and fused 9/7 transcode kernels on the existing CUDA C PTX path for a later focused slice
• update env var and unsafe audit docs plus strict GPU parity coverage

Validation

Local:

• cargo fmt --check
• cargo fmt --check --manifest-path crates/j2k-cuda-runtime/src/cuda_oxide_transcode/simt/Cargo.toml
• git diff --check
• cargo test -p j2k-cuda-runtime --all-targets --features cuda-oxide-transcode
• cargo clippy -p j2k-cuda-runtime --all-targets --features cuda-oxide-transcode -- -D warnings
• cargo test -p xtask --test repo_lint
• cargo xtask unsafe-audit

CUDA host jcwal@100.75.125.59:

• cargo oxide build --arch sm_89
• strict transcode runtime with J2K_REQUIRE_CUDA_OXIDE_TRANSCODE=1 and J2K_CUDA_USE_OXIDE_TRANSCODE=1
• strict combined cuda-oxide runtime for copy, encode, decode-store, dequantize, idwt, and transcode
• cargo +nightly clippy --manifest-path crates/j2k-cuda-runtime/src/cuda_oxide_transcode/simt/Cargo.toml -- -D warnings