Torchtitan CI gap Between ROCm & CUDA

[akashveramd]

Creating this issue to fill torchtitan CI gap between ROCm & CUDA.
We need to add support for integration tests & tests under experiments. First priority is for integration tests. The integration tests include features.py, flux.py, ft.py, h100.py and models.py tests. Experiments contain following tests compiler_toolkit, simple_fsdp, torchcomms, transformers_modeling_backend and vlm tests.

• Implement ciflow/rocm for torchtitan. This will allow ROCm workflows to run from forked PRs.

- Integration Tests-

• Currently, ROCm only supports features integration test. This PR added ROCm CI support for features tests Enable ROCm CI support #1786. For both CUDA & ROCm, the execution time for features integration test is ~18 minutes. Pull Docker step takes ~5 minutes. Hence, the total execution time for features integration test is ~23 minutes.
• models integration test
• flux integration test
• ft integration test
• h100 integration test

- Experiment Tests-

• compiler_toolkit experiment test
• simple_fsdp, torchcomms experiment test
• transformers_modeling_backend experiment test
• vlm experiment test

There are CPU Unit tests. However, they don't need ROCm support as these are CPU tests than runs on linux.2xlarge runner.