dmdAnal: Ensure that ExnStr flag isn't dropped inappropriately

This fixes #13977 and consequently #13615. Here an optimization in the
demand analyser was too liberal, causing us to drop the ExnStr flag and
consequently resulting in incorrect demand signatures. This manifested
as a segmentation fault in #13615 as we incorrectly concluded that an
application of catchRetry# would bottom.

Specifically, we had

    orElse' :: STM a -> STM a -> STM a
    orElse' x = catchRetry# x y
      where y = {- some action -}

Where the catchRetry# primop places a demand of <xC(S),1*C1(U)> on its
first argument. However, due to #13977 the demand analyser would assign
a demand of <C(S),1*C1(U)> on the first argument of orElse'. Note the
missing `x`.

    case orElse' bottomingAction anotherAction of { x -> Just x }

being transformed to,

    case orElse' bottomingAction anotherAction of {}

by the simplifier. This would naturally blow up when orElse' returned at
runtime, causing the segmentation fault described in #13615.

Test Plan: Validate, perhaps add a testcase

Reviewers: austin, simonpj

Reviewed By: simonpj

Subscribers: rwbarton, thomie

GHC Trac Issues: #13977, #13615

Differential Revision: https://phabricator.haskell.org/D3756