Preserve progn-like clauses for coverage
Removing spurious progns makes for a cleaner macro-expansion. However,
because SBCL maps cons cells in the original source file to source paths
when recording coverage instrumentation, discarding cons cells
unnecessarily leads to worse coverage instrumentation.
Generally, we can't record form coverage for atoms, but if the atom is
in a progn-like form, we can get accurate coverage information by
mapping the cons cell whose car contains that atom to its source-path.
This is handled in maybe-instrument-progn-like in
src/compiler/ir1-tran.lisp.