ftrace,bug: Improve traceoff_on_warn

ftrace,bug: Improve traceoff_on_warn

While doing some tracing, I found a huge portion of the per-cpu buffer
was taken by printk/serial output because we're disabling the trace far
too late (after printing the CUT string).

Improve matters for architectures that have GENERIC_BUG + _BUG_FLAGS by
killing the tracer in the exception handler before printing anything
much.

Link: https://lkml.kernel.org/r/20200528145240.GF706495@hirez.programming.kicks-ass.net
Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org>
Signed-off-by: Steven Rostedt (VMware) <rostedt@goodmis.org>