ARROW-11683: [R] Support dplyr::mutate()

[nealrichardson]

First steps:

• Rework selected_columns to hold field_refs instead of string column names; add code to back out the string field names where needed (e.g. dataset Project())
• Create an array_ref pseudo-function to do the same as field_ref for array_expressions
• Add a data argument to eval_array_expression in order to bind array_refs to the actual Arrays before evaluating
• Refactor filter() NSE code for reuse in mutate()
• Split up dplyr tests because we're going to be adding lots more

Then:

• Basic mutate() and transmute() (done in 578d492)
• Go through the examples in the dplyr::mutate() docs and add tests for all cases. Where possible they're implemented in arrow fully; where we don't support the functions, it falls back to the current behavior of pulling the data into R first.

Followup JIRAs:

• ARROW-11704: Wire up dplyr::mutate() for datasets
• ARROW-16999: Implement dplyr::across() and autosplicing
• ARROW-11700: Internationalize error handling in tidy eval
• ARROW-11701: Implement dplyr::relocate()
• ARROW-11702: Enable ungrouped aggregations in non-Dataset expressions
• ARROW-11658: Handle mutate/rename inside group_by
• ARROW-11705: Support scalar value recycling in RecordBatch/Table$create()
• ARROW-11754: Support dplyr::compute()
• ARROW-11752: Replace usage of testthat::expect_is()
• ARROW-11755: Add tests from dplyr/test-mutate.r
• ARROW-11785: Fallback when filtering Table with if_any() expression fails

[github-actions]

https://issues.apache.org/jira/browse/ARROW-11683

[jonkeane]

This looks good, I went through commit by commit which helped before going back over the full diff. A few suggestions and a few questions (making sure I am reading this correctly)