nv50/ir: do not perform global membar for shared memory

nv50/ir: do not perform global membar for shared memory

Shared memory is local to CTA, thus we should only wait for
prior memory writes which are visible to other threads in
the same CTA, and not at global level. This should speedup
compute shaders which use shared memory.

Signed-off-by: Samuel Pitoiset <samuel.pitoiset@gmail.com>
Reviewed-by: Ilia Mirkin <imirkin@alum.mit.edu>