[vulkan] Shaders caching (#39384)
Summary:
Pull Request resolved: https://github.com/pytorch/pytorch/pull/39384
Introducing `ComputeUnitFactory` which is responsible for providing `ComputeUnit`s (Shaders),
it caches it, using shader name (glsl file name)+workGroupSize as a cacheKey, just `std::map<string, std::shared_ptr>`
Macro GLSL_SPV changed to have literal name for cache key as a first argument.
All constructors of ComputeUnit are changed to use `ComputeUnitFactory`
Ownership model:
ComputeUnitFactory also owns `vkPipelineCache` that is internal vulkan cache object ( https://www.khronos.org/registry/vulkan/specs/1.2-extensions/man/html/VkPipelineCache.html )
`VContext` (global object) owns ComputeUnitFactory, that owns ComputeUnits, vkPipelineCache, for destruction of them we need valid VkDevice, so it should be destructed before `vkDestryDevice` in `~VContext` => As members of the class will be destructed only after destructor - forcing destruction of ComputeUnitFactory before `vkDestroyDevice`, doing `unique_ptr<ComputeUnitFactory>.reset()`
Test Plan: Imported from OSS
Reviewed By: AshkanAliabadi
Differential Revision: D21962430
Pulled By: IvanKobzarev
fbshipit-source-id: effe60538308805f317c11448b31dbcf670487e8
