Do not use VLAs in vec256_qint.h (#36855)
Summary:
Use `std::decay_t<decltype(foo)>::size()` instead of `foo.size()` to help compiler with static array allocations.
Even if `Vec256::size()` is `constexpr`, `foo.size()` (where `foo` is of type `Vec256`) is not an integral constant expression, therefore compiler have to use VLAs, which are not part of C++ standard.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/36855
Test Plan: CI
Differential Revision: D21151194
Pulled By: malfet
fbshipit-source-id: eaf3e467c7f7ee6798ca82fe9f8fae725011ead0
