Add positions to all expression and pattern nodes. Kill off the LExpr
and LPattern expression and pattern cases that held just a position.
The idea had been to always use the first enclosing such position.
This doesn't provide accurate enough positions for type errors (in
general for a type error you always want the exact expression you're
complaining about, especially if you're printing spans of positions
and not just line numbers), and, worse, with this approach it's hard
to tell if any given position is the one you actually ought to be
using.

Fix up some of the position usage where prompted by the ensuing typing
changes.

Kill off the currentExprPos field of the local state monad in MGU.hs,
which was for collecting the enclosing LExpr and is no longer needed.

In pursuit of all this, add two ops to Position.hs:
   - leadingPos, which gets the empty (no span) position at the beginning
     of the position of something else, which can be used for deriving the
     position of implicit elements;
   - maxSpan', which is like maxSpan but works for (exactly) two Positioned
     objects of potentially heterogeneous type. This allows dropping a
     number of explicit calls to getPos in places where that's verbose and
     ugly.

These contain source positions, so derived Eq instances aren't in
general going to produce the desired behavior.

Leave the Eq instance on Type for the moment because it's used in the
typechecker. This is going to require further attention.

There were only two uses of it; provide matching-based alternatives
for both, on the grounds that if we don't need an explicit Eq instance
that specifically ignores position information it's better not to have
to write one.

The use in MGU handled peeling off any LType nodes; the one in
Interpreter did not and was therefore potentially wrong...

We concluded after discussion that using the more precise position
from the pattern involved is preferable to using PosREPL, even though
the name "it" is predicated on this coming from the repl.

Also add a bunch of infrastructure -- the driver script for this is
massively overkill for just this test but we expect this test to grow
a lot of siblings in the not-too-distant future.