[ATen][rocm] fix type issue (#72676)
Summary:
Pull Request resolved: https://github.com/pytorch/pytorch/pull/72676
Fixes a nightly build failure at AMD.
Originally,
```
std::initializer_list<int64_t> size = { param_size * num_linear_layers / 2, 1};
```
but https://github.com/pytorch/pytorch/pull/71925 changes as follows.
```
const auto size = { static_cast<int64_t>(bias_size * num_linear_layers / 2), 1L};
```
This caused a problem as follows (with a certain compiler):
```
/pytorch/aten/src/ATen/native/miopen/RNN_miopen.cpp:356:108: error: no matching function for call to 'at::Tensor::set_(const c10::Storage&, size_t&, const std::initializer_list<const long int>&)'
```
This patch reverts the type of `size`.
```
std::initializer_list<int64_t> size = { static_cast<int64_t>(bias_size * num_linear_layers / 2), 1L};
```
This issue happens with GCC <= 9.1, but not with Clang and GCC >= 9.2. See this: https://godbolt.org/z/sbErY7her
(Note: please let me explcitly cast the first variable (`bias_size * num_linear_layers / 2`) to avoid signed/unsigned type conversion warning. It should not cause any issue.)
Test Plan: CI
Reviewed By: malfet
Differential Revision: D34153890
fbshipit-source-id: 791434c57ed669f95a36a79584bad5586ef66a28
(cherry picked from commit e6435065544da05087004253d856c5ce8269b2af)
