Revert to intrinsic memcpy instead of inline

PYTHON-450
intrinsic was 25% faster in benchmarks

cassandra/cython_marshal.pyx

@@ -18,6 +18,7 @@ import six
 
 from libc.stdint cimport (int8_t, int16_t, int32_t, int64_t,
                           uint8_t, uint16_t, uint32_t, uint64_t)
+from libc.string cimport memcpy
 from cassandra.buffer cimport Buffer, buf_read, to_bytes
 
 cdef bint is_little_endian
@@ -25,15 +26,6 @@ from cassandra.util import is_little_endian
 
 cdef bint PY3 = six.PY3
 
-cdef inline void memcopy(char *dst, char *src, Py_ssize_t size):
-    """
-    Our own simple memcopy which can be inlined. This is useful because our data types
-    are only a few bytes.
-    """
-    cdef Py_ssize_t i
-    for i in range(size):
-        dst[i] = src[i]
-
 ctypedef fused num_t:
     int64_t
     int32_t
@@ -59,7 +51,7 @@ cdef inline num_t unpack_num(Buffer *buf, num_t *dummy=NULL): # dummy pointer be
         for i in range(sizeof(num_t)):
             out[sizeof(num_t) - i - 1] = src[i]
     else:
-        memcopy(out, src, sizeof(num_t))
+        memcpy(out, src, sizeof(num_t))
 
     return ret
 

cassandra/numpy_parser.pyx

@@ -153,7 +153,7 @@ cdef inline int unpack_row(
             Py_INCREF(val)
             (<PyObject **> arr.buf_ptr)[0] = <PyObject *> val
         else:
-            memcopy(<char *> arr.buf_ptr, buf.ptr, buf.size)
+            memcpy(<char *> arr.buf_ptr, buf.ptr, buf.size)
 
         # Update the pointer into the array for the next time
         arrays[i].buf_ptr += arr.stride