[compiler toolkit] Add graph based SAC

[yiming0416]

Add graph-based SAC as a joint graph pass in alternative to eager-mode AC. Previously eager AC and fullmodel trace don't compose in DSv3 (See #1935). With graph based AC, it's no longer an issue.

Set --compile.joint_passes apply_sac to turn it on

Llama3-8B (TP=2, FSDP=2)
no-AC: memory: 67.19GiB(37.67%) tps: 9,569 tflops: 554.17 mfu: 24.63%
eager SAC: memory: 45.88GiB(25.73%) tps: 8,627 tflops: 499.62 mfu: 22.21%
graph SAC: memory: 44.20GiB(24.78%) tps: 8,007 tflops: 463.74 mfu: 20.61%

DeepSeek-v3 Debug model(TP=2, FSDP=2, EP=2)
no-AC: memory: memory: 12.26GiB(6.88%) tps: 59,927 tflops: 32.54 mfu: 1.45%
eager SAC: doesn't work (due to AC HOP mutation issue #1935)
graph SAC: memory: 11.01GiB(6.17%) tps: 53,939 tflops: 29.29 mfu: 1.30%

[aditvenk]

How will user pass the op_list_to_save from command line?

Can we re-use the existing AC config system to configure apply_sac_pass?

[yiming0416]

Currently the ops to be saved in SAC are in a hardcoded list for each model.
Not sure if it's a good UX for user to pass from command line though, as the list itself could be pretty long and not cli-friendly