feat: implement long adder to optimize the counter of new metrics system

[empiredan]

https://github.com/apache/incubator-pegasus/issues/889

[Smityz]

LGTM

[Smityz]

Can you construct a benchmark for its performance compared with the origin perf_counter?

[empiredan]

Can you construct a benchmark for its performance compared with the origin perf_counter?

Actually the benchmark given in https://github.com/apache/incubator-pegasus/issues/889 has involved the comparison with the original perf-counter: divided_long_adder is nearly the same with the original dsn::perf_counter_number_atomic except that its virtual functions are removed since other long adders also have no virtual function.

The implementation of divided_long_adder can be found in src/utils/long_adder_bench/long_adder_bench.cpp.

[acelyc111]

What's th difference between this PR's implemention and the original https://github.com/apache/kudu/blob/master/src/kudu/util/striped64.h?

[empiredan]

What's th difference between this PR's implemention and the original https://github.com/apache/kudu/blob/master/src/kudu/util/striped64.h?

The data structures and algorithms between both striped long adder are based on striped64. The differences are:

• This PR defines a default constructor for the atomic cache-line-aligned int64, while kudu defines a non-trivial one.
• This PR encapsulates the "create function" of the atomic cache-line-aligned int64 as a common utility which is also exposed and can be used by external modules.
• This PR introduces unique_ptr<...> to manage the memory of cells, while kudu releases memory by calling free() in the destructor.
• This PR supports fetch_and_reset, since we have volatile scenarios where only recent count will be showed.
• This PR fixes some styles, for example, replace if (c) with if (c != nullptr).
• This PR uses dsn::utils::get_current_tid() to get the initial hash code, while kudu uses Random to get.
• In "internal reset", this PR only sets "base" with target value, with cells set to zero, while kudu set all of "base" and cells to target value.
• Besides striped_long_adder, this PR also implements concurrent_long_adder to optimize the scenario that the adder is updated very frequently.
• This PR uses long_adder_wrapper to wrap the interfaces of long adder.