use spin lock in auto growth allocator

[wanghuancoder]

PR types

Others

PR changes

Others

Describe

测试发现，在多线程申请显存、释放显存过程中，会发生锁碰撞。此时，会有1个线程进入休眠状态。当该线程获得锁后，再从休眠状态唤醒。这个过程是非常耗时的。直接影响了执行器的调度性能。
为此，我们将std::mutex 换成 自旋锁，自旋锁在发现锁已经被lock时，本线程会自旋等待，不进入休眠状态，实测性能比std::mutex好很多。
本PR只修改了auto growth allocator，其它的allocator不清楚都是什么依赖关系，暂未修改。

[paddle-bot-old]

Thanks for your contribution!
Please wait for the result of CI firstly. See Paddle CI Manual for details.

[sneaxiy]

Some comments are listed as follows. @zhiqiu @wanghuancoder

1. Can you show any testing example and data?
2. Linus has addressed that: spinlocks can only be used if you actually know you're not being scheduled while using them. See here.

[wanghuancoder]

Some comments are listed as follows. @zhiqiu @wanghuancoder

1. Can you show any testing example and data?
2. Linus has addressed that: spinlocks can only be used if you actually know you're not being scheduled while using them. See here.

我是在为新版本执行器写GC时发现的该问题，我用文字描述一下我的测试：
有2根线程，1根线程负责mutabledata+LaunchKernel。另1跟线程使用cudaEventSynchronize等待特定OP的kernel执行结束，则释放相应Tensor的显存。测试模型是PTB，BatchSize=20。由于PTB都是小kernel，因此会高频的mutabledata和freeAllocation。统计发现，7655806次加锁中（申请释放显存都会加锁），会有904919次发生锁碰撞（碰撞，就是1跟线程加锁，另一根线程申请锁）。这种情况下，整个测试用时36秒，其中mutable_data用时9.27秒，freeAllocation用时9.96秒。

如果换成1跟线程同时负责以上2跟线程的工作（即不发生锁碰撞），整个测试用时24秒，mutable_data用时2.212，reeAllocation用时1.557秒。

如果仍是2根线程，将Allocator的锁改成自旋锁，整个测试用时27秒，mutable_data用时3.83，reeAllocation用时6.72秒。

[sneaxiy]

LGTM for some suggestions.

• Please add some notes about where the codes of spin_lock.h is from.
• Try to implement a BasicLockable class instead of using macro everywhere.

class SpinLock {
  public:
    void lock();
    void unlock();
    DISABLE_COPY_AND_ASSIGN(SpinLock);
};