[vulkan] Remove the persistent resource pool (#66478)
Summary:
Pull Request resolved: https://github.com/pytorch/pytorch/pull/66478
A persistent resource pool was needed to store prepacked tensors since the main resource pool tied to the global Vulkan context would be flushed at the end of each inference run. However, prepacked tensors needed to alive between inference runs, so an additional persistent resource pool was introduced that would only be flushed when the Vulkan context was destroyed.
However, with [this change](https://github.com/pytorch/pytorch/pull/66477) the resource pool no longer indiscrimately flushes allocated resources at the end of an inference run. Tensors will have to call `release_resources()` before they become eligible to be destroyed. Since prepacked tensors are tied to an `OpContext` object they will stay alive between inference runs.
Therefore, the persistent resource pool is no longer needed.
Test Plan: Build and run `vulkan_api_test`.
Reviewed By: beback4u
Differential Revision: D31490076
fbshipit-source-id: 3741a2333c834796d589774e819eaaf52bb9f0fe