Add cache for parseNodeName

[chunnienc]

This PR adds cache for parseNodeName - an expensive string manipulation function in graph executor. The time saved is around 2% to 3% for MobileNetV3 (tested on my laptop, may be imprecise).

To see the logs from the Cloud Build CI, please join either our discussion or announcement mailing list.

---

This change is 

[mattsoulanille]

LGTM. I was going to recommend using a closure to memoize parseNodeName instead of adding it to context but it looks like you need to pass this context to other functions like getParamValue.

[chunnienc]

Yeah. I think the context is suitable since it covers calls from parseNodeName itself and also getTensor and getParamValue, and it associates the lifecycle of the strings with the graph, so that these string keys can be gc'ed with the graph.

[Linchenn]

Thanks! I think this approach would benefit small models a lot, by introducing cache and cache lookup to avoid string operations.

Have you benchmarked some large models? For large models, the number of nodes may be significant, so I am not sure, when the cache is very large, whether the cache lookup is still faster than string operations?

[chunnienc]

Thanks! I think this approach would benefit small models a lot, by introducing cache and cache lookup to avoid string operations.

Have you benchmarked some large models? For large models, the number of nodes may be significant, so I am not sure, when the cache is very large, whether the cache lookup is still faster than string operations?

(Discussed offline) Yes it's always faster.
Benchmark: https://jsbench.me/8eleeumdpi/1

[Linchenn]

LGTM, thanks!