[TESTING] use torch.int for autotuning cache (#840)

For stupid reasons, ops on int8 are 3 times slower than on int, and for
another set of stupid reasons we are not using cudaMemset for `zero_`,
so using `int8` buffer in `do_bench` makes it slow.

Co-authored-by: Philippe Tillet <phil@openai.com>

python/test/regression/test_performance.py

@@ -128,7 +128,7 @@ elementwise_data = {
         1024 * 16: 0.0219,
         1024 * 64: 0.0791,
         1024 * 256: 0.243,
-        1024 * 1024: 0.534,
+        1024 * 1024: 0.530,
         1024 * 4096: 0.796,
         1024 * 16384: 0.905,
         1024 * 65536: 0.939,