Implement breakpoint() using PTX brkpt instruction

See https://docs.nvidia.com/cuda/parallel-thread-execution/#miscellaneous-instructions-brkpt

This is implemented using the low-level API because an overload with the
high-level API causes the breakpoint to look like a function call that
the code has entered, which is a bit confusing for the user.

numba_cuda/numba/cuda/cudadecl.py

@@ -508,6 +508,11 @@ def generic(self, args, kws):
             return signature(dty, ary, idx, dty, dty)
 
 
+@register_global(breakpoint)
+class Cuda_breakpoint(ConcreteTemplate):
+    cases = [signature(types.none)]
+
+
 @register
 class Cuda_nanosleep(ConcreteTemplate):
     key = cuda.nanosleep

numba_cuda/numba/cuda/cudaimpl.py

@@ -934,6 +934,14 @@ def ptx_atomic_cas(context, builder, sig, args):
 
 # -----------------------------------------------------------------------------
 
+
+@lower(breakpoint)
+def ptx_brkpt(context, builder, sig, args):
+    brkpt = ir.InlineAsm(ir.FunctionType(ir.VoidType(), []),
+                         "brkpt;", '', side_effect=True)
+    builder.call(brkpt, ())
+
+
 @lower(stubs.nanosleep, types.uint32)
 def ptx_nanosleep(context, builder, sig, args):
     nanosleep = ir.InlineAsm(ir.FunctionType(ir.VoidType(), [ir.IntType(32)]),