Migrate TODO.md to issues...

These should go into github issues
# TODO
- Add a TupleSize variant to ArraySizeEstimate to deal with Scanl'.
(Or just get rid of Scanl/r'!!)
- FreeVars should somehow be attached to specific Lambdas, not to
array ops.
- types like Type, Prim, Const should probably be reexported by GPUIR
and CLike IRs so that those modules are self contained.
- GPUIR needs an overhaul. Kernel declaration probably need to be
separate from kernel invocation (e.g. a fold will be multiple
kernel launches), and kernels need to have local work size as well
as global.
- The interpreters are in disarray. It would be nice to have a fully
functional interpreter for all three major IRs.
- UnzipETups: really needs to use maybeLet
- (BUG) p20\* was exposing a bug where a singleton ETuple would appear
out of OneDimensionalize.
- FuseGenReduce: inline Cond's as well.
- InlineCheap should inline Use dereferences.
- Perhaps factor TArray into its own type. It really shouldn't be
mixed up with the scalar types, this just encourages incomplete
pattern matches.
- Why don't we use Word instead of Int by default for all shape
related stuff?
- Fix sizeAnalysis to work for the replicates in Nbody. Is it not
seeing through "EShape"?
- Compile EIndexScalar of known index into Use by just doing the
dereference...
## DONE
- Need to track the FULL SHAPE of any progResults for reconstructing
the final thing. (Fix OneDimensionalize & downstream)
-- ?? Is this still relevant? [2013.03.03]
-- No, WithShapes should address [2014.02.24]