* Update ROCm instructions for manylinux

* Sync devtoolset version to 11 with CUDA manylinux_2_28 entry

* Update Dockerfile for ROCm

* Add almalinux case

* Use | to OR different conditions and avoid repetition

* Use devtoolset 11 in alignment with CUDA manylinux config

* Magma build preinstalls

(cherry picked from commit 9a803c9)

* Add symlink to python

(cherry picked from commit 5c4bf9f)

* Add ROCM_PATH

* Use cpu_final to get the devtoolset PATH/LD_LIBRARY_PATH settings and pip cmake install

* Move creation of python link to install_rocm_magma.sh since it's only needed for Magma

* Cmake updates

* Add Almalinux case for bundling system libs

* Set _GLIBCXX_USE_CXX11_ABI=1 for ROCm since RHEL8 ROCm packages refer to cxx11-abi symbols, so PyTorch build fails if _GLIBCXX_USE_CXX11_ABI=0

* Need * to have a match

* Install numpy<2.0.0 for older release branches

* Enable MIOpen build from source for ROCm6.1 to see if it resolves segfault issue
Remove existing sqlite3.pc to make cget happy

* Add conditions for ROCm6.2 and ROCm6.3

* Don't use set -e

* Upgrade cmake version to enable CK to build for ROCm6.2

* Disable building from source for MIOpen and reinstate set -e

* Add librocm-core.so (#50)

(cherry picked from commit 2d3fb6f)

* Remove all references to devtoolset

* Update cmake version to 3.28.4

* Skip epel-7 rpm download

* Skip epel-7 rpm download

* Install epel-7 from PyTorch AWS backup repo (similar to upstream PR pytorch/pytorch#132449

* Use DEVTOOLSET_VERSION 9 for manylinux2_28

* Move AOTriton build step to PyTorch build script (since different aotriton for different pytorch releases)

* Install libaotriton_v2.so before we look for it

* Comments

* Comments

* Revert "Remove all references to devtoolset"

This reverts commit 43d70a7.

* Comments

* Add gfx942 as default arch

* Use correct GPU_IMAGE for manylinux_2_28

* Tab lint

* Do not rebuild CK; Remove old ROCm cruft

---------

Co-authored-by: Pruthvi Madugundu <pruthvigithub@gmail.com>