Consider try / catch functionality

Currently `error "foo"` is like throwing an exception, but there is no way to catch the exception.

Motivations for "catch":
1) Routines that loop over data and trigger errors for particular items don't have useful stack traces because the line of code in the stack trace doesn't tell you e.g. which array index caused the problem. This is particularly important in custom manifestation functions where we typically iterate over the tree forcing fields.
2) It is currently not possible to write test cases in Jsonnet that check that the correct error is raised.
3) It is not possible to detect using reflection that a field requires overriding, because we typically represent that with `f: error "Must overide f"` see #414.

Caveats:
- Should it be possible to catch "runtime errors" like `1/0` or `[1,2,3][10]`? If not, how do we distinguish?
- Should it be possible to catch major errors like exceeding stack size? If not, how do we distinguish?
- It is clearly not possible to catch non-termination (halting problem).
- Should error values be cached in thunks?