Add option to support garbage collection after torch compilation (#2559)

Summary:

X-link: pytorch/pytorch#142821

This diff is an extension of ezyang's PR https://fburl.com/6uvvzb4f.
In ezyang's PR above, it adds gc after torch compilation finished. 
The gc operation is guarded by jk: pytorch/compiler:enable_run_gc_after_compile
The gc op time cost will be logged into dynamo_compile scuba table.

This diff extends the PR to: 
  - Use garbage collection on Generation 1 instead of generation 2 (default), which greatly reduced the gc latency overhead from 160 sec per rank to 10 sec per rank.
  - Additionally introduce an environment variance which has the higher priority than the JK to control whether we do gc or not after the torch compilation. (default value set to gc enabled). This environment variance will be used for AB testing of training jobs to compare the pt2 compilation time and memory cost.

Reviewed By: ezyang, yuxihu

Differential Revision: D67062158

Author: qiurc

userbenchmark/dynamo/dynamobench/_dynamo/utils.py

@@ -882,6 +882,7 @@ class CompilationMetrics:
     triton_version: Optional[str] = None
     feature_usage: Optional[dict[str, bool]] = None
     compile_time_autotune_time_us: Optional[int] = None
+    gc_time_us: Optional[int] = None
 
 
 DEFAULT_COMPILATION_METRICS_LIMIT = 64