[refactor] Get rid of unnecessary snodes map in Program class.

[ailzhang]

Found this issue while reading snode tree related code :P
It mainly does two things:

1. Get rid of snode map in program class, we seem to use it in async engine and llvm cpu backend but it's not necessary.
2. Somehow we compile struct on both device and host when arch is cuda (in LlvmProgramImpl::materialize_snode_tree), it turns out to be redundant as well.

[netlify]

✔️ Deploy Preview for jovial-fermat-aa59dc canceled.

🔨 Explore the source changes: a4cf4d4

🔍 Inspect the deploy log: https://app.netlify.com/sites/jovial-fermat-aa59dc/deploys/616fe2b22501d000080193c8

[k-ye]

LGTM!