Add option to disable cubin cache binaries and moved some redundant kernel load code into specific function

src/Makefile

@@ -7,6 +7,7 @@ SHARED                  := 0
 DEBUG                   := 0
 PRODUCTION              := 0
 PRODUCTION_VERSION      := v5.1.0
+ENABLE_CUBIN            := 1
 ENABLE_BRAIN            := 1
 USE_SYSTEM_LZMA         := 0
 USE_SYSTEM_ZLIB         := 0
@@ -234,6 +235,11 @@ LFLAGS                  += -lxxhash
 endif
 endif
 
+# CUDA binary cache
+ifeq ($(ENABLE_CUBIN),1)
+CFLAGS                  += -DWITH_CUBIN
+endif
+
 ##
 ## Native compilation target
 ##