ci: split the test archive by crate

[MarcusKainth]

Stacked on #584 (ci/archive-no-debuginfo); the diff against that branch is this one commit.

What this changes, and why

Every test job downloaded one archive holding every crate's test binaries, and build-tests built that archive and the ROM suites archive one after the other. A simulation group needs only the native crate's suites.

• build-tests is a matrix of two archives built side by side: native.tar.zst (the native crate's suites) and workspace.tar.zst (every other crate's). Each leg has its own rust-cache key.
• The ROM suites archive moves to build-refemu, which already builds the reference emulator in release and isn't on the path to the test jobs.
• Each group downloads only what scripts/test-group.sh runs it from: the simulation groups get the native archive, emulator gets the workspace and ROM suites archives, and native-rest gets both. It runs the native crate's other suites and the driver's native_* suites as two nextest invocations, one per archive, still at two threads each. native_connections_live stays in emulator at --test-threads 1.
• A group that lacks an archive it needs fails on the missing file.

ci-passed still needs build-tests, which now reports the result of both legs together.

Evidence

Each group selects the same tests

Every group's filter over the whole workspace before the split, against the split filters over the two archives after (nextest list --message-format oneline, sorted, compared with cmp), with local archives built the way CI builds them:

emulator       old  457 new  457 same
native-sim-a   old   10 new   10 same
native-sim-b   old   12 new   12 same
native-sim-c   old   12 new   12 same
native-sim-d   old   12 new   12 same
native-sim-e   old   28 new   28 same
native-sim-f   old   21 new   21 same
native-rest    old  401 new  401 same
every test in the workspace:      953; in some group after the split:      953; duplicates:        0
exit=0

This PR's CI run 36124399424

Each build leg has a new rust-cache key, so this run compiled the dependencies too. The warm figures for the same split come from the measurement run 36122853256, with the native crate at opt-level 0 (#585): 48 s for the native archive and 94 s for the rest.

	#584 alone (run 36122982586, attempt 2, warm)	this PR (cold cache)
build-tests	1 job, 342 s: 316 s archive, 6 s upload	2 jobs side by side: native 229 s (199 s archive), workspace 275 s (248 s archive)
ROM suites archive	in build-tests	build-refemu, 36 s. The job takes 67 s and gates only emulator
artifacts	test-archives 191,650,427 B	native-tests 101,225,013 B, workspace-tests 70,556,324 B, rom-suites 5,064,897 B
download per group	3 to 7 s, both archives	3 to 6 s, only the group's own archives

The nextest summaries show every group running what it ran before:

test (emulator)       Summary [ 196.129s] 457 tests run: 457 passed, 0 skipped
                      Summary [  29.100s] 10 tests run: 10 passed, 0 skipped
test (native-rest)    Summary [  69.286s] 382 tests run: 382 passed, 0 skipped
                      Summary [ 143.356s] 19 tests run: 19 passed, 0 skipped
test (native-sim-a)   Summary [  75.511s] 10 tests run: 10 passed, 0 skipped
test (native-sim-b)   Summary [ 160.567s] 12 tests run: 12 passed, 0 skipped
test (native-sim-c)   Summary [ 107.962s] 12 tests run: 12 passed, 0 skipped
test (native-sim-d)   Summary [ 145.380s] 12 tests run: 12 passed, 0 skipped
test (native-sim-e)   Summary [ 163.015s] 28 tests run: 28 passed, 0 skipped
test (native-sim-f)   Summary [ 224.610s] 21 tests run: 21 passed, 0 skipped

The whole run took 9 min 15 s from the first job's start (10:31:02) to ci-passed (10:40:17). native-rest ran in 220 s as two invocations, against 356 s in #584's attempt 2 and 276 s in #584's attempt 1.

$ make actionlint zizmor shellcheck; echo "exit=$?"
No findings to report. Good job! (9 suppressed)
exit=0

Invariants

None. CI layout only; the same tests run, from different archives.

Spec impact

• None. No contract in SPEC.md is touched

Checks

• make gates: not run locally. CI's lint job ran make lint and every test group passed
• No AI attribution trailers in the commits

Written mostly by Claude Opus 5.5.