ci(publish): sync publish.yml with all staging fixes

Port all proven fixes from publish_staging.yml to publish.yml for main releases:

- Split monolithic matrix build into per-platform jobs (Linux, Linux-CUDA, macOS, Windows)
- Linux/Linux-CUDA: use Docker container (coasys/ad4m-ci-linux) with pre-installed deps
- Linux: set PKG_CONFIG_PATH explicitly for container jobs (soup3-sys/webkit fix)
- Linux-CUDA: add CUDA_COMPUTE_CAP=80 env var
- macOS: pin macos-14 runner, Xcode 15.4 (ld-classic for branch islands),
  CARGO_NET_GIT_FETCH_WITH_CLI (chromium submodule fix), Rust/Homebrew caching
- Windows: generate-deno-snapshot on Linux + download artifact (CRT mismatch workaround),
  NASM hiding (aws-lc-sys prebuilt fix), RUSTFLAGS +crt-static, long paths,
  manual JS dep build order, Rust caching
- All platforms: RUSTY_V8_MIRROR for pre-built V8 binaries
- npm-publish: switch to direct npm publish commands, add executor package
- crates-publish: use Docker container, add 'already exists' handling
- Add generate-deno-snapshot job (shared Linux→Windows artifact)

Author: marvin-bot-coasys