improve span accuracy during parsing

[JeanMertz]

Once #6023 is merged, VRL will have its first batch of diagnostic messages.

While the initial batch is "good enough", there's more we can do.

Take there two examples:

error: unneeded error assignment
  ┌─ :2:1
  │
2 │ ok, err = 5;
  │ ^^^^^^^   - because this expression cannot fail
  │ │          
  │ this error assignment is unneeded
  │
  = hint: assign to "ok", without assigning to "err"
  = see language documentation at: https://...

error: unknown function argument keyword
  ┌─ :2:11
  │
2 │ to_string(foo: .bar, .baz)
  │           ^^^^^^^^^^^^^^^ unknown keyword: foo
  │
  = see language documentation at: https://...

While they are useable, the span ranges aren't as specific as I'd like.

For example, in the first one, I'd rather have the span point to err specifically. And for the second one it should point to the foo keyword.

This is mostly a matter of slicing up our parse functions into smaller sections that only parse one specific part, and then encode the span of that part in their result. We're close to that goal, but this issue can track the progress to resolving the above to cases, and any others where more fine-grained span information would improve the diagnostic.