Memory planning debug log (#2125)

Summary:
Calculate `num_reuse_pairs` only if debug is enabled
Reduces to_executorch latency.

thanks mergennachin for the finding!

Cprof before:
{F1459766592}

Cprof after:
{F1461507195}

Pull Request resolved: https://github.com/pytorch/executorch/pull/2125

Test Plan:
ci

Latency (s) for language llama fp32:

```
python -m examples.models.llama2.export_llama -c ../llama-models/language/very_new_checkpoint.pt -p ../llama-models/language/params.json --fairseq2 --dtype-override fp32 -prof /tmp/language-fp32-memop.html
```
||to_executorch|memory_planning_pass|
|before|58.7s|14.7s|
|after|42.4s| 0.466s|

Reviewed By: mergennachin

Differential Revision: D54214624

Pulled By: lucylq

fbshipit-source-id: 367862704176a514f09a3e5c41e85d9e2019c8e9