When loading a large number of modules into GHCi, we collect
CompiledByteCode for every module and then link it all at the end.
Space leaks in the CompiledByteCode linger until we traverse it all for
linking, and possibly longer, if there are bits we don't look at.

This is the nuke-it-from-orbit approach: we deepseq the whole thing
after code generation. It's the only way to be sure.

Test Plan:
Heap profile of GHCi while loading nofib/real/anna into GHCi, this patch
reduces the peak heap usage from ~100M to ~50M.

Reviewers: hvr, austin, bgamari, erikd

Subscribers: thomie

Differential Revision: https://phabricator.haskell.org/D2419