Contain xcodebuild's derived data in the test's temp dir

[erneestoc]

Every xcodebuild test-without-building invocation creates a ~/Library/Developer/Xcode/DerivedData/temporary-<hash>/ directory that nothing ever deletes. When the runner does not pass -resultBundlePath — i.e. any xcodebuild-path run without create_xcresult_bundle — that directory receives a full .xcresult of the run; and passing -resultBundlePath alone does not fix it, a bookkeeping directory still leaks per run.

One development machine had accumulated 267 of these directories totaling 26GB. Every CI agent running simulator tests through the xcodebuild path leaks one per test run, silently.

Fix: pass -derivedDataPath pointed into the runner's test_tmp_dir, which the existing EXIT trap already deletes. An explicit -resultBundlePath (from create_xcresult_bundle or xcodebuild_args) still wins for the result bundle itself.

Measurements

New DerivedData temporary-* directories created per invocation (Xcode 26.2, iOS 26.2 simulator, hosted unit test):

xcodebuild flags	new dirs per run	contents
(none — today's runner without create_xcresult_bundle)	1	152K, including a complete orphan .xcresult
-resultBundlePath only	1	24K bookkeeping
-derivedDataPath (this PR)	0	—
both	0	—

End-to-end through the runner after this change: default mode, dir count unchanged across runs (was +1 per run); with create_xcresult_bundle = True, the bundle still lands in TEST_UNDECLARED_OUTPUTS_DIR (verified present in outputs.zip) and DerivedData stays untouched.

Steps to reproduce the leak

ls -d ~/Library/Developer/Xcode/DerivedData/temporary-* | wc -l
# run any hosted test through ios_xctestrun_runner without create_xcresult_bundle,
# e.g. a unit test with a test_host (forces the xcodebuild path):
bazel test //your:HostedTest --runner=@rules_apple//apple/testing/default_runner:ios_xctestrun_ordered_runner --nocache_test_results
ls -d ~/Library/Developer/Xcode/DerivedData/temporary-* | wc -l   # +1, forever

Test plan

• Manual xcodebuild test-without-building flag matrix above (each variant run against the same xctestrun).
• Runner end-to-end: default mode creates no DerivedData directories; CREATE_XCRESULT_BUNDLE=1 still produces tests.xcresult in undeclared outputs with DerivedData untouched.
• Rendered template passes bash -n.

[erneestoc]

The CI failures here are the pre-existing wedged-simulator flake, not this change:

• The failing set (HelloWorldSwiftUITests, ButtonsUITests, ios_coverage_ui_test_coverage_new_runner) is the same session-sensitive group that has been failing on other PRs whose runner-test caches get invalidated (Give each concurrent test action its own simulator via a per-(device, OS) pool #3026 has the full diagnosis: a previously-terminated test leaves the reused, still-booted simulator with a dead device-side session, and every later xcodebuild test-without-building against it hangs or fails to initialize).
• The attempt log shows the signature exactly: reused booted BAZEL_TEST_iPhone 16_26.0, a well-formed invocation with -derivedDataPath, then silence until the timeout SIGTERM.
• 1163 tests passed on the same run — including the hosted unit tests, which exercise -derivedDataPath on every invocation.
• All three failing targets pass locally on this branch (9.8s / 11.6s / 5.3s).

#3026 adds the recovery machinery for this (its own CI is now fully green, including the previously-failing 9.x job, because its first run self-heals a wedged agent). Suggest retrying this build after #3026 lands — no change needed here.

[adincebic]

Good.