Cache Rust builds and remove ineffective thin LTO

[rob-p]

Summary

• cache Rust dependency artifacts in the regular Linux/macOS CI matrix
• enable cargo-dist's supported per-target release-build cache and regenerate the 0.31 workflow
• remove thin LTO from release and distribution builds
• remove simpleaf's unused direct cc dependency (cc remains transitively where native builds need it)
• move checkout, Rust setup, and setup-miniconda to their supported Node 24-era actions

The cache deliberately leaves the simpleaf crate itself uncached. Rust-cache keys dependency artifacts by runner, toolchain, Cargo manifests/lockfile, and compiler environment. Release caching is declared through cache-builds in dist-workspace.toml, rather than as a hand edit to generated CI.

LTO measurement

Controlled clean-target release builds used the same checkout, host, Rust toolchain, and 16-way build:

configuration	wall	peak RSS	unstripped binary	stripped xz
thin LTO	101.89 s	4,142,812 KiB	128,445,728 B	15,726,188 B
LTO off	95.58 s	3,643,372 KiB	129,739,680 B	15,541,484 B

Disabling thin LTO reduced wall time by 6.2% and peak compiler RSS by 12.1%. The distributable-size proxy was 1.2% smaller without LTO.

Five alternating 1,000-process batches showed no measurable runtime change: median quant-help time was 1.84 s for both binaries, and median chemistry-registry load/lookup time was 1.80 s for both. Because simpleaf delegates mapping, collation, quantification, and index construction to child programs, there is no demonstrated end-user runtime benefit to pay for thin LTO here.

CI cache measurement

An identical rerun of the final commit, after its cache-seeding run completed, produced:

runner	cold	warm	reduction
ubuntu-latest	16m25s	2m49s	82.8%
macos-15	13m01s	2m21s	81.9%

Both warm jobs ran the full workflow, including release build/tests, strict rustdoc on Linux, conda environment setup, the released alevin-fry artifact, and the toy end-to-end pipeline. The cache archives were 858 MB on Linux and 516 MB on macOS.

Validation

• all 77 release-profile tests pass
• strict rustdoc passes
• Cargo metadata resolves with the lockfile
• both workflows parse as YAML
• cargo-dist 0.31 regenerates the release workflow cleanly and generate --check passes