[Feature Request]: Explore and implement robust sandboxing mechanism

### Feature description

Currently `eval_code` evaluates each call in the new (cloned) environment where `parent.env(env) <- parent.env(.GlobalEnv)`. When executing `library` via `eval_code` new `library` is added over each `qenv@env` - i.e. all `qenv` objects have the same package-namespaces in the search path.

We would like to separate `@env` as much as possible.

### Separate search path for each `env@env`

So that each `qenv` is a child of libraries which have been attached in this `eval_code`


### Separate `qenv` sessions

Other solution is to separate sessions where `eval_code` is executed [rlang::r_session](https://callr.r-lib.org/reference/r_session.html)

Start by tracing how eval_code clones environments and how qenv@env shares package namespaces. Compare separate search paths with the proposed rlang::r_session approach; done means eval_code executions have isolated environments and package namespaces under a clearly selected design.

Scritto dal modello di indicizzazione a partire dal testo della issue.