ci(build): split allTests into parallel per-target jobs to end OOM flakes

[michalharakal]

Problem

build-job intermittently fails after ~15 min with "The operation was canceled" and no BUILD FAILED — the flaky ~20-minute reds (e.g. PRs #850, #852). Investigation shows this is runner memory exhaustion, not a code defect:

• The job log goes silent minutes before the cancel (runner becomes unresponsive), then GitHub reaps it — the classic system-OOM signature. No OutOfMemoryError, no failing test.
• build.yml ran assemble allTests in a single ubuntu-latest (7 GB) job. allTests fans out the Gradle JVM (-Xmx4g) + Kotlin daemon (1 GB) + per-target test forks + Node (JS/Wasm) + the native C toolchain concurrently; peak RSS sometimes tips over 7 GB → OOM-kill. Sometimes it just fits (which is why ci(docs): publish on release and stop fork PRs failing on comment #850 went green on retry). It passes locally because dev machines have more RAM.
• Bonus: the Memory info diagnostic used if: failure(), but an OOM-killed runner reports cancelled, not failed — so it never captured data on exactly these runs.

Fix

• Split the test run into a 3-leg matrix — jvm / js-wasm / native — each running one target family (jvmTest / jsTest wasmJsTest wasmWasiTest / linuxX64Test). No single job holds the whole footprint, and the legs run in parallel, so wall-clock is the slowest leg instead of the sum. Directly targets both the OOM flakiness and the ~20-min duration.
• assemble becomes its own memory-light job, preserving the build/package guarantee for all targets (incl. those without tests).
• Gradle caching per job (same pattern/pin as docs.yml) so parallel legs don't cold-download the dependency graph.
• Memory diagnostic now fires on cancelled() too, so future OOMs are actually captured.
• Aggregate build-job gate keeps the exact same check name, so the existing required status check keeps working with no branch-protection change.

Coverage / verification

• The three legs cover every runnable Linux test target that allTests ran (Apple targets are macOS-only and were already skipped on Linux).
• Task names verified: linuxX64 is a declared target; jvmTest jsTest wasmJsTest wasmWasiTest dry-run resolves cleanly.
• allTests passes locally (BUILD SUCCESSFUL); no production code changed here — workflow only.

Follow-up

If the js-wasm leg ever approaches the limit, it can be split further (js vs wasm) — and the memory snapshot will now be captured to confirm.