add @ as py attr accessor

[t-kalinowski]

Closes #1429
Closes #251

This enables unambiguously accessing attrs in contexts where $ mingles python attrs and items (e.g., with python dictionaries, pyspark tables). This also enables autocompletion in R outside the terminal. However, autocompletion in the IDE won't work until rstudio/rstudio#13451 is resolved.

[dfalbel]

Maybe we can add docs in this PR explaining the differences between $, [, [[ and @ so we can also fix #263.

[t-kalinowski]

I don't think merging this at this time is a good idea.

@ is valuable new syntax, which doesn't come around very often. Using it for python attribute access today doesn't strike me as a good return on value--it wouldn't add any new capability over what $ does already in practice. I don't have any better ideas for how to use @ in reticulate at this moment, but I'd rather reserve the syntax for a better idea, when it does come along.

Inside a package that uses reticulate, (e.g., pysparklyr), I think it's reasonable for package authors to opt-in to a "safer" attribute access method that doesn't fallback to __getitem__ by defining an internal S3 method, e.g.,

`@.python.builtin.object` <- function(x, name) py_to_r(py_get_attr(x, name))