tvm operator dynolog (#26295)
Summary:
Pull Request resolved: https://github.com/pytorch/pytorch/pull/26295
Log the following in scuba caffe2_tvm_operator_stats:
1. everything in caffe2_operator_stats
2. fallback netdef
3. tvm module graph_json
4. whether compilation triggered this round
5. number of compilations stored in tvm_runtime_map
6. (not yet logged) last compilation time if any
7. (not yet logged) total bytes occupied by compilation
8. whether this compilation is fallback
9. batch size as observed by tvm op
Test Plan:
```
buck run mode/dbg //tvm/sparse:tvm_bbpredictor_benchmark -- --init_net ~/tmp/ads/84480054_204/init_net.pb --input_init_net ~/tmp/ads/84480054_204/input_init_net.pb --pred_net ~/tmp/ads/84480054_204/pred_net.pb --warmup 1000 --iter 1000 --num_cycles 5 --caffe2_logging_operator_dyno_sampling_rate=1 --vmodule=Logger=
2
```
Logs show up in the scuba:
https://our.intern.facebook.com/intern/scuba/query/?dataset=caffe2_tvm_operator_stats
https://fburl.com/scuba/lq2h22e4
Auto submitted adindexer canary:
https://our.intern.facebook.com/intern/ads/canary/421064436039494716
Additional adindexer canary:
https://our.intern.facebook.com/intern/ads/canary/421082681202831286/
Additional adfinder canary:
https://our.intern.facebook.com/intern/ads/canary/421082685084831037/
Reviewed By: yinghai
Differential Revision: D17358412
fbshipit-source-id: d2119c12ddeaa86217c163e32fb1e211952139f5
