Merge failures due to stats increase

[Bryan Richter]

Last night, Marge's merge requests failed several times because of a
stats increase.

compile_time/bytes allocated increased from
x86_64-linux-fedora33-release baseline @
def47dd32491311289bff26230b664c895f178cc:
    Expected    T9872a (normal) compile_time/bytes allocated: 1655471856.0 +/-1%
    Lower bound T9872a (normal) compile_time/bytes allocated:   1638917137
    Upper bound T9872a (normal) compile_time/bytes allocated:   1672026575
    Actual      T9872a (normal) compile_time/bytes allocated:   1674083264
    Deviation   T9872a (normal) compile_time/bytes allocated:          1.1 %
*** unexpected stat test failure for T9872a(normal)

This happened on three separate attempts, each with a deviation of 1.1%:

* https://gitlab.haskell.org/ghc/ghc/-/jobs/1247176
* https://gitlab.haskell.org/ghc/ghc/-/jobs/1247312
* https://gitlab.haskell.org/ghc/ghc/-/jobs/1247512

The reason I'm making an email out of this is that none of the
individual patches that Marge is merging have this problem, but it
seems to happen consistently when they are combined.

You can see a list of patches by looking at the list of mentions on
https://gitlab.haskell.org/ghc/ghc/-/merge_requests/9439.

There were other problems with Marge's merges last night that I DO
understand and have fixed. But this one, of course, is over my head.
:) Maybe somebody can look at the list of patches and deduce something
useful. Otherwise, I will continue monitoring Marge's progress, and
try to find a smaller set of patches that can make it through. I'll
prioritize getting the Javascript backend through.

-Bryan