This is a regression introduced in e1112ef, which was causing
TestNilInterfaceError in ./tests package to fail.

This is another regression introduced in e1112ef, this error is
actually used in prelude's $assertType function, along with a fake
`_type` struct.

Admittedly, this isn't the best fix, but reconciling type information
with the reflect types is a rabbit hole I'm unwilling to jump into right
now.

A new test case involving unsafe.Pointer was added, which we have to
skip, since GopherJS doesn't support raw pointers. However, I also added
rudimentary support for calling `reflect.Value.Convert()` with
`unsafe.Pointer` to make sure the test doesn't panic.

This test relies on assumptions about string and slice implementation
that do not in fact hold true for GopherJS.

Due to an incorrectly passing type assertion the test triggers a code
path that panics instead of returning an error, which causes the test to
fail. In the grand scheme of things there's nothing wrong with the io
package though.

GopherJS runtime doesn't provide access to low-level details such as
alignment, and atomics are emulated anyway.

This applies to both GOROOT and GOPHERJS_GOROOT. The motivation behind
the change is that setting a variable to an empty value is a common way
of "unsetting" it and most tools treat it that way.