KASSERT: always assert; KWARN

[Conrad Meyer]

I'd like to logically revert r243980 and r244105, such that KASSERT
uses the __dead2-annotated panic(9).

Going back to the old behavior enables Coverity and other static
analyzers to reason about KASSERT invariants via the __dead2 panic(9)
path.

This proposal is in https://reviews.freebsd.org/D6117 .

As a follow-up, to match the assumed intent of the r243980 changes, I
propose a KWARN facility which may be muted, rate limited, or even
cause panic.  Generally, KASSERTs should not be KWARNs.  That proposal
is here: https://reviews.freebsd.org/D6134

Finally, I am looking for suggestions of things it *does* make sense
to KWARN about.  One suggestion was witness_warn; however, it doesn't
seem like a great fit (without adding allocating sbufs in, anyway).  A
sketch of that is in https://reviews.freebsd.org/D6306 .

Thoughts or objections?  Does anyone like the ability to opt out of
invariants asserts?

Best,
Conrad