[RFC] Consolidate simple_fsdp and compiler_toolkit experiments

[yiming0416]

This PR merges the simple_fsdp and compiler_toolkit experiments into a new unified experiment called graph_based_training (name to be discussed later).

The two experiments shared the same DTensor-based SimpleFSDP model authoring but had separate compilation paths: simple_fsdp used JIT compilation (torch.compile) and compiler_toolkit used AOT joint graph capture. The new experiment unifies them under a single compile.mode config field ("jit" or "aot"), with a shared pass registry that validates pass/mode compatibility.

This PR creates a separate new folder. No existing files in simple_fsdp/ or compiler_toolkit/ are modified.

File change breakdown

Files copied without changes:

• simple_fsdp.py — Copied from simple_fsdp/simple_fsdp.py.
• reshard_after_forward.py — Copied from simple_fsdp/reshard_after_forward.py.
• cudagraph.py — Copied from compiler_toolkit/cudagraph.py.

Files copied with import path changes only

• common_utils.py — Adapted from compiler_toolkit/common_utils.py.
• graph_utils.py — Adapted from compiler_toolkit/graph_utils.py.
• jit_backend.py — Adapted from simple_fsdp/backend.py.
• train.py — Adapted from compiler_toolkit/train.py.
• llama3/__init__.py — Adapted from simple_fsdp/llama3/__init__.py.
• llama3/model.py — Adapted from simple_fsdp/llama3/model.py.
• deepseek_v3/__init__.py — Adapted from simple_fsdp/deepseek_v3/__init__.py.
• deepseek_v3/model.py — Adapted from simple_fsdp/deepseek_v3/model.py.

Files adapted with non-trivial changes

• passes.py — Unified pass registry
• compilation.py — Unified compilation dispatcher routing to _apply_jit() and _apply_aot()
• job_config.py — Merged config with mode aot, jit
• llama3/parallelize.py — Unified parallelize function merging logic from simple_fsdp and compiler_toolkit
• deepseek_v3/parallelize.py — Same as above, but for DSv3.
• tests/integration_tests.py — Merged integration tests from both experiments.

[fegin]

Can we use git mv to preserve the history if a file is copied from the original folder? We can deprecate the two experiments in this PR. With git mv it is more clear to see what is going on with this change.

[tianyu-l]

sorry I'm doing a massive refactoring of torchtitan config system. I'm almost done (with the first version) and I'd prefer we rebase after I'm done.. (pray)

[yiming0416]

moved to #2457