Using eg, fsck with the MD5 backend loads whole files into memory.

May only happen for very large files (40 gb) or in other specific
circumstances, including ghc version used for buildd etc.

Observed with 6.20160217-g95bbdb8, linux standalone amd64.

Not observed with 5.20151218-g5008846.

Commit 7482853ddddc21f2696dcfbc82d737f03032134a may be relevant,
but I don't understand how yet. A small test program using the same
code doesn't exhibit the problem, even when built in the identical build
environment as the 6.20160217-g95bbdb8 that has the problem.

> Update: Reverted 7482853ddddc21f2696dcfbc82d737f03032134a and indeed the
> problem got fixed. But, reverting that commit breaks the test suite on
> windows and has a FD leak, so is not desirable. This needs more
> investigation. --[[Joey]]

>> I see it now, the checksum is a String and it was only forced to WHNF,
>> so the hashing didn't fully complete and the file got buffered.
>> Probably only occurred when fscking, and not when adding a file,
>> due to differing use patterns of the checksum.
>> [[fixed|done]] --[[Joey]]