[cuda] Port over allocator and buffer implementation

This commit ports over existing CUDA driver allocator and buffer
implementation. The main logic is kept as-is, with one noticeable
changes--context wrapper is dropped and fields in it are directly
put in various API calls. This is to make supporting multiple
device and stream easier later. Other changes are just polishing
on comments and errors.

experimental/cuda2/CMakeLists.txt

@@ -17,6 +17,10 @@ iree_cc_library(
     "api.h"
   SRCS
     "api.h"
+    "cuda_allocator.c"
+    "cuda_allocator.h"
+    "cuda_buffer.c"
+    "cuda_buffer.h"
     "cuda_driver.c"
   DEPS
     ::dynamic_symbols