Qualcomm AI Engine Direct - Enable zero copy feature (#2531)

Summary:
- Add argument "shared_buffer" into compiler_spec, qnn_executor_runner and test scripts
  -  Actually, shared_buffer should be a runtime option since user are responsible to allocate memory for tensors on device. But it seems to have no way to set the runtime option to QnnBackend. Therefore, we put it to compile_spec for now.
- Implement SharedBuffer to allocate and free RPC memory
- Add QnnMemManger to register shared buffer for tensor
  - During exection time, we will register memory of tensor data for QNN. And we will deregister them during destruction time of QnnBackend
- Add two API `void* QnnExecuTorchAllocCustomMem(size_t bytes, size_t alignment)` and `void QnnExecuTorchFreeCustomMem(void* buffer_ptr)` to allocate RPC memory with SharedBuffer
  -  Users are responsible to allocate "enough" tensor bytes, and set alignment as MemoryAllocator::kDefaultAlignment. See runtime/core/memory_allocator.h.

Pull Request resolved: https://github.com/pytorch/executorch/pull/2531

Reviewed By: kirklandsign

Differential Revision: D55142607

Pulled By: cccclai

fbshipit-source-id: 6a394ef0023e70362c3cf963a0519c3efaa95bc2