Test for performance of rewriting plugins seems flaky

The TcPlugin_RewritePerf test, which measures the performance impact of rewriting plugins on type-family reductions, seems to exhibit large decreases in allocations in some CI runs (98% decreases):

compile_time/bytes allocated decreased from x86_64-linux-deb9-hadrian baseline @ a975583c70e57434340d9a20c976c8f06fde9beb:
    Expected    RewritePerfDefs.hs (normal) compile_time/bytes allocated: 2239480832.0 +/-1%
    Lower bound RewritePerfDefs.hs (normal) compile_time/bytes allocated:   2217086023 
    Upper bound RewritePerfDefs.hs (normal) compile_time/bytes allocated:   2261875641 
    Actual      RewritePerfDefs.hs (normal) compile_time/bytes allocated:     47791056 
    Deviation   RewritePerfDefs.hs (normal) compile_time/bytes allocated:        -97.9 %

The issue seems to be that we are not always measuring the same thing in this test: sometimes all the modules in the test are tested separately, and sometimes they are lumped together.