[RocksDB] Expose thread local perf counter for low overhead, per call level performance statistics.

Summary:
As title. No locking/atomic is needed due to thread local. There is also no need to modify the existing client interface, in order to expose related counters.

perf_context_test shows a simple example of retrieving the number of user key comparison done for each put and get call. More counters could be added later.

Sample output
./perf_context_test 1000000
==== Test PerfContextTest.KeyComparisonCount
Inserting 1000000 key/value pairs
...
total user key comparison get: 43446523
total user key comparison put: 8017877
max user key comparison get: 88939
avg user key comparison get:43

Basically, the current skiplist does well on average, but could perform poorly in extreme cases.

Test Plan: run perf_context_test <total number of entries to put/get>

Reviewers: dhruba

Differential Revision: https://reviews.facebook.net/D12225

include/leveldb/perf_context.h

@@ -0,0 +1,24 @@
+#ifndef STORAGE_LEVELDB_INCLUDE_PERF_CONTEXT_H
+#define STORAGE_LEVELDB_INCLUDE_PERF_CONTEXT_H
+
+#include <stdint.h>
+
+namespace leveldb {
+
+// A thread local context for gathering performance counter efficiently
+// and transparently.
+
+struct PerfContext {
+
+  void Reset(); // reset all performance counters to zero
+
+
+  uint64_t user_key_comparison_count; // total number of user key comparisons
+};
+
+extern __thread PerfContext perf_context;
+
+}
+
+
+#endif