ci: build MCU SoC timing IR on Linux + de-flake opensta-to-ir locate test

[robtaylor]

Summary

Unblocks PR #63's CI by fixing two unrelated failures it surfaced:

1. MCU SoC Metal Simulation — the second cosim invocation was passing --sdf / --sdf-corner, which cosim never accepted. Historically masked because clap rejected the now-renamed --max-cycles first. Cosim's documented contract is "pre-convert with opensta-to-ir and pass --timing-ir <.jtir>"; raw SDF support is the docs/plans/post-phase-0-roadmap.md follow-up.

   New prepare-mcu-soc-jtir Linux job builds OpenSTA, installs the pinned sky130A PDK via volare, runs opensta-to-ir against tests/mcu_soc/data/6_final.sdf, and uploads the resulting 6_final.jtir as an artifact. The Metal job needs: it, downloads the artifact, and passes --timing-ir to the second cosim. Removes the redundant "Strip SDF timing checks" step (was overwriting the tracked stripped SDF with identical content).

2. opensta-to-ir locate test flake — locate_reports_failing_version_probe occasionally got VersionProbeFailed { kind: ExecutableFileBusy } instead of the expected VersionProbeNonZero. Linux's close(2) does not synchronously decrement the inode writer_count, so an immediate execve of the freshly-written stub races. Added sync_all plus a bounded no-op execve probe loop in the test helper so the stub is provably ready before the test runs.

Also documents the Python tooling convention (uv dev dependencies; volare/ciel for PDK install) in CLAUDE.md.

Test plan

• prepare-mcu-soc-jtir job completes: OpenSTA + opensta-to-ir build, volare installs sky130A, jtir artifact uploaded
• MCU SoC Metal Simulation job downloads the jtir and completes "Capture stimulus + timing VCD via cosim" without CLI parse errors
• opensta-to-ir Tests job: locate_reports_failing_version_probe passes (validating the de-flake)
• Downstream mcu-soc-cvc / mcu-soc-comparison jobs see expected stimulus.vcd / timing VCD artifacts

Notes for PR #63 rebase

PR #63's rename --max-cycles → --max-clock-edges still applies cleanly: the first cosim (line 626) is untouched here. The second cosim (was line 653) is rewritten to use --timing-ir; PR #63's rename of that line will need to be re-applied to the new --max-cycles 10000 \ line. set -o pipefail additions are not in this PR's scope.

Followups still open (tracked in docs/plans/post-phase-0-roadmap.md):

• Add --sdf subprocess wrapper to cosim so the Linux jtir prep job can be inlined
• Migrate volare dep to upstream-renamed ciel package