[5/N][SAC] Refactor activation checkpointing to use centralized policy-based approach, and recompute lora params

[mori360]

The previous activation checkpointing implementation had each model define its own _op_sac_save_list in its respective parallelize.py file. Manually maintaining the op list risks missing standard compute-intensive ops that PyTorch's get_default_op_list() already provides.

This PR:

• Centralize ops in activation_checkpoint.py via default_activation_checkpoint_policy()
• Use get_default_op_list() from torch._functorch.partitioners as the foundation
• Split ops list into compute and communicate case, and extend them with the current ops used in torchtitan but not included in get_default_op_list
• Specially wrap policy by ac_config.per_op_sac_force_recompute_mm_shapes_by_fqns
• Recompute small params and skip them in count for every other param save policy.

Stack from ghstack (oldest at bottom):

• -> [5/N][SAC] Refactor activation checkpointing to use centralized policy-based approach, and recompute lora params #2489
• [4/N] Add QAT (quantization-aware training) model converter #2488
• [3/N] Support qlora by nf4tensor #2487
• [2/N] Support lora checkpoint on partial save and multi-source load #2485
• [1/N] Add lora adapters for finetune #2484