Another attempt to deflake arm64 tests

[bduffany]

• Reduce the effective --local_test_jobs from 16 to 8 - seems like remote_execution_test keeps crashing due to running too many jobs locally. We're already setting size = "enormous" so I'm not sure there's anything else we can do to reduce the local test concurrency, short of setting tags = ["exclusive"] which we don't really want since the individual test shards take kind of a long time :/
• Increase flaky_test_attempts from 2 to 3
• Remove cpu:4 tag - I found this on stackoverflow or something a while back, but when testing it out it doesn't seem like it actually does anything in this case. In the timing profile, I was seeing 16 test actions running concurrently which wouldn't make sense if this actually worked, since the arm64 runner only has 16 cores.

Related issues: N/A