Ensure on.exit() runs at end of chunk

[hadley]

Fixes #201

@cderv this has the potential to affect existing Rmds, but I think it's low risk as currently on.exit() runs immediately making it effectively useless, so I can't see why anyone would have called it.

[hadley]

This also changes the behaviour of return(). Previously:

evaluate::evaluate(c(
  "1 + 1",
  "return()", 
  "2 + 2"
))
#> <evaluation>
#> Source code: 
#>   1 + 1
#> Text output: 
#>   [1] 2
#> Source code: 
#>   return()
#> Text output: 
#>   NULL
#> Source code: 
#>   2 + 2
#> Text output: 
#>   [1] 4

Now:

evaluate::evaluate(c(
  "1 + 1",
  "return()", 
  "2 + 2"
))
#> <evaluation>
#> Source code: 
#>   1 + 1
#> Text output: 
#>   [1] 2
#> Source code: 
#>   return()

That seems like better behaviour to me, but it also has the possibility of affecting existing code.

[hadley]

Also worth noting we'll need to update withr to match this behaviour.

[hadley]

Just noticed that with this PR,

```{r}
#| error: true
mpg
```

returns

mpg
#> Error in eval(as.call(list(cb)), envir): object 'mpg' not found

[lionel-]

Could we zap all error calls where the expression is eval(...)?

[hadley]

@lionel- we're already doing that and it works fine within evaluate itself, but it some how leaks out when used in knitr.

[hadley]

@lionel- now fixed; I don't understand why stop("1") and 3() would get different calls, but maybe you do?

[cderv]

I agree with the change, it makes on.exit useful, and I don't think it was for now.