release OCaml's main lock when computing `Sha1.file_fast`

[c-cube]

The computation is potentially costly, and since it's in C it should be parallelizable. I looked and it doesn't seem to be the case right now.