Fix thread-local PooledStackAllocator init order

[beef9999]

use_pooled_stack_allocator可能在程序很早的地方被调用，然而，PooledStackAllocator是一个thread-local变量，由于是按需使用alloc/dealloc，所以这个变量的初始化也很晚。

在程序结束的时候，PooledStackAllocator会优先于其他thread-local模块析构，于是造成了这些thread-local模块（已经使用了它分配的内存）产生coredump。

这个MR在use_pooled_stack_allocator的时候就初始化thread-local变量，并且把 set_thread_stack_allocator 放到 thread.h 模块，简化一些 stack-allocator.h 的代码

---

use_pooled_stack_allocator may be called very early in the program. However, PooledStackAllocator is a thread-local variable, and due to the on-demand use of alloc/dealloc, the initialization of this variable is also very late.​
At the end of the program, PooledStackAllocator will be destructed before other thread-local modules, which causes some thread-local modules (that have used the memory allocated by it) to generate coredumps.​
This MR initializes the thread-local variable when use_pooled_stack_allocator is called, and moves set_thread_stack_allocator to the thread.h module to simplify some code in stack-allocator.h.

[Coldwings]

use_pooled_stack_allocator是需要严格要求首次photon init是配置的（或者在vcpu_init之前进行配置），并且不允许改回。因此这个配置的处理才是单向的（仅能开启，不能通过设置为false进行关闭）。

对于明确知道栈分配器配置的用户场景，使用单独的photon::set_thread_stack_allocator可以自由控制更改；对于非此类用户，应当避免在使用默认option去初始化主线程外的vcpu时，由于使用默认配置而改回，这样会造成pool泄漏、释放不正确等问题。

[Coldwings]

each vcpu will call photon_init, and use_pooled_stack_allocator is a global config.
This change may leads to user set-up pooled-stack-allocator and then unset. (by initialize another vcpu without setting use_pooled_stack_allocator. That is highly risky.

[Coldwings]

Since it is a one-way setter and always put same value, is it really need to use a std::call_once?

[Coldwings]

The only problem is here:
once vcpu_init called in a global constructor, it might be called before photon_thread_alloc construct or library load.

that may cause delegate photon_thread_alloc become an empty delegate and unable to be called, result in allocation failure and segment fault.