IsString rules

[Bodigrim]

Closes #140 in line with design suggested at #140 (comment)

The idea is that users were never meant to use IsString.fromString directly, this method should be invoked only by {-# LANGUAGE OverloadedStrings #-} magic. This PR effectively enforces this limitation.

• In ghci IsString.fromString = packCharsSafe. This is a partial function,
  throwing errors on characters out of Latin1 range, but partiality is not a big
  deal for interactive sessions.
• For compiled programs:
  • For ASCII literals IsString.fromString = packChars = unsafePackLiteral,
    no runtime overhead and Core remains exactly the same as before.
  • For Unicode literals (and literals with '\0') IsString.fromString = packCharsSafe.
    This incurs runtime overhead, but such thunk is evaluated only once in its life.
  • For dynamic strings (user input, etc.) IsString.fromString = error.
    This is to make runtime behaviour more robust (if it crashes, it crashes for
    any input) and avoids a danger of routines, which were tested only for ASCII
    inputs but suddenly receive Unicode and silently lose data.

[sjakobi]

The main problem that I'm expecting with this approach is that programs compiled with -O0 or -fno-enable-rewrite-rules may behave in unexpected ways. For example, test suite runs with -O0 might fail.

[Bodigrim]

I kinda lost my plot here: CI fails, and I don't really remember what, why and how. We might be able to find better approaches in future, so in the meantime closing.