incorrect gradient

added DisconnectedType

helper function

None

ambiguous whether to turn numpy into sparse or dense variable

the old exception message was annoyingly uninformative
corrected mathematically inaccurate comment

IndexError if it gets the wrong number of gradient terms from an Op

we still support this behavior but want to forbid it soon

supported now

made grad_sources_inputs accept None for inputs

types

variables to marks

marked some tests as KnownFailure--they fail due to unconventional
Op.grad implementation

implemented an unimplemented gradient case, this got rid of most of the
KnownFailureTests this branch introduces

still run it to compare against new version of grad)

tests harder

just one way

fixed bug; cost was considered a disconnected input