ci: cache the Lean toolchain (or retry elan) — TLS download failures are failing runs #1426

In the last ~2 hours, 7 Lean Action CI runs failed before any Lean step ran. In each, leanprover/lean-action@v1 couldn't download the toolchain:

error: could not download file from 'https://releases.lean-lang.org/lean4/v4.36.0-rc1/lean-4.36.0-rc1-linux.tar.zst'
info: caused by: [35] SSL connect error (TLS connect error ...)

Affected runs include 40641067857, 40640554181, 40640421642, 40640072473, 40639256732, 40638743720 and 40637480920. Each one needs a manual rerun. A failure inside a merge-group run would eject every PR in that group.

Proposal: stop depending on a live toolchain download for every job.

• Cache ~/.elan (toolchains) with actions/cache, keyed on lean-toolchain, before lean-action runs (or use whatever toolchain-cache option lean-action offers).
• Alternatively, or as well, retry the elan install a few times with backoff.

Keep the existing Lake/Mathlib cache behaviour unchanged.

Done when: a warm run doesn't contact releases.lean-lang.org for the toolchain, and a cold-cache run retries transient download failures before failing.