Adds missing restores of the stolen register on translation paths where the target thread is not in the code cache: at various system call or stub locations.
Tested by manually enabling the tool.drcacheoff.burst_threadfilter test (it was disabled for AArch64 because of the #2007 (closed) link failure in some toolchains) and confirming the crash disappears.
Enables more tests on AArch64: client.drwrap-test-detach, api.static_noclient, api.thread_churn.
Fixes #4468 (closed)