storage: TestRaftRemoveRace is flaky under stress (ambiguous result)

[tamird]

Somewhat hard to reproduce locally, but reproduces almost instantly in GCE:

tamird@gceworker173:~/go/src/github.com/cockroachdb/cockroach$ make stress STRESSFLAGS='-p 100 -stderr -maxfails 1' PKG=./pkg/storage TESTS=TestRaftRemoveRace
go list -tags '' -f 'go test -v  -tags '\'''\'' -ldflags '\'''\'' -i -c {{.ImportPath}} -o {{.Dir}}/stress.test && (cd {{.Dir}} && if [ -f stress.test ]; then stress -p 100 -stderr -maxfails 1 ./stress.test -test.run '\''TestRaftRemoveRace'\'' -test.timeout 2m ; fi)' ./pkg/storage | /bin/bash
github.com/cockroachdb/cockroach/pkg/storage
github.com/cockroachdb/cockroach/pkg/storage_test
testmain
0 runs so far, 0 failures, over 5s
0 runs so far, 0 failures, over 10s
0 runs so far, 0 failures, over 15s
2 runs so far, 0 failures, over 20s
9 runs so far, 0 failures, over 25s
57 runs so far, 0 failures, over 30s
96 runs so far, 0 failures, over 35s
98 runs so far, 0 failures, over 40s

--- FAIL: TestRaftRemoveRace (7.48s)
	client_test.go:915: change replicas of range 1 failed: result is ambiguous (Raft re-proposal failed: write at timestamp 0.000000123,1564 too old; wrote at 0.000000123,1565)
FAIL


ERROR: exit status 1

99 runs completed, 1 failures, over 41s
FAIL
Makefile:138: recipe for target 'stress' failed
make: *** [stress] Error 1

My guess is that we need better error handling in either (*storage.Replica).ChangeReplicas or its callers (*storage.multiTestContext).{,un}replicateRange.