This cleans up the package subsystem a little.  There are some
changes to the GHC API as a result.

  - GHC.init and GHC.initFromArgs are no longer necessary.

  - GHC.newSession takes the root of the GHC tree as an argument
    (previously passed to GHC.init).

  - You *must* do GHC.setSessionDynFlags after GHC.newSession,
    this is what loads the package database.

  - Several global vars removed from SysTools

  - The :set command in GHCi can now cause new packages to be loaded,
    or can hide/ignore existing packages.

To catch the case of not running darcs-all get.

This commit does two largely-unrelated things, but they hit the same code.

First, I tweaked the error messages a bit, to give better errors
for impredicative polymorphism.  This added the mb_fun argument to
tc_sub.

Second, I fixed a long-standing bug in tc_sub.  In the isBoxyTyVar case 
of tc_sub (rule F2) I was not recursing to tc_sub as the rule suggests,
but rather calling u_tys.  This is plain wrong, because the first
arugment might have more foralls.   

The solution is to recurse to tc_sub, but that in turn requires a parameter,
exp_ib, which says when we are inside a box.

Test is tc210.

This is a long-standing bug really (Trac #900).  The poly_id passed
to tcSpecPrag should be zonked, else it calls tcSubExp with a non-zonked
type; but that contradicts the latter's invariant.

I ended up doing a bit of refactoring too.  The extra lines are 
comments I think; the code line count is reduced.

Test is tc212.hs

Instantiate data constructor type variables with the type constructor type variables in constraints for derived class instances. Fixes instance deriving for GADTs which declar H98 types, closing ticket #902.

Fixed error in user's manual, gadt section, part about deriving. Data constructor type was Maybe, should be Maybe1.

	MERGE TO 6.6 branch!

Some while ago I made the type checker a tiny bit more lenient about
left sections, so that 
	(x !)
would typecheck iff
	((!) x) 
typechecks.  

Strictly, Haskell 98 requires that the section typechecks iff
	(\y. (!) x y)
typechecks, and I should really have made the relaxation dependent on a 
flag, but I didn't.

Anyway, this commit fixes the *desugarer* so that it correctly desugars
the programs that the typechecker passes.

When fiddling with pattern-matching for unboxed tuples, I'd messed up
the slightly-tricky tests for pattern matching on unboxed tuples, notably
	case (# foo, bar #) of r -> ...r...

The fix is in TcPat, and test are tcfail115, tcfail120, and tc209