Fix import chaining for custom header support - #618

Fixes dhall-lang/dhall-json#60

In the above issue, a remote import specified custom headers using a
relative path:

```
$ curl https://gist.githubusercontent.com/reactormonk/a8192fe784608f1f7515d5583c448095/raw/d46f5b9aea9eb56c4bdd92d187936ee87e9f4cf2/local.dhall

let test = https://gist.githubusercontent.com/reactormonk/a8192fe784608f1f7515d5583c448095/raw/e8e5b14c2afe9acc342fe111bcf6e09b90dced9c/test.dhall using ./headers in test
```

But import chaining was not working correctly for the `using` clause of the
import, meaning that before this change the `./headers` import was not
anchored to the parent import.

Using the `</>` judgment from the standard, the previous behavior was:

```
http://example0.com/foo </> http://example1.com/bar using ./baz

= http://example1.com/bar using ./baz
```

... when the behavior should have been:

```
http://example0.com/foo </> http://example1.com/bar using ./baz

= http://example1.com/bar using http://example0.com/baz
```

Note that custom header support has not yet been fully standardized, but the
latter version is what the correct behavior should be once it is standardized.

The reason the former version is problematic is because it does not behave
correctly in the presence of header forwarding.  If `http://example1.com/bar`
imports any remote expressions of its own, it will use the relative import
`./baz` to customize the headers for downstream imports from the same domain.
However, because this relative import is not anchored to the parent import, it
will change for each downstream import.

For example, suppose that the contents of `http://example1.com/bar using ./baz` are:

```
http://example1.com/baz
```

Import chaining that after its parent import would give:

```
http://example1.com/bar using ./baz </> http://example1.com/baz

= http://example1.com/baz using ./baz
```

... which introduces an inadvertent import cycle because now `./baz` will
resolve to `http://example1.com/baz` instead of `http://example0.com/baz`.  An
import cycle similar to this one caused the bug described by the linked issue.

This change fixes import chaining to correctly anchor custom header imports to
the parent import and this fix resolves the above issue.

This also includes a small change to fix canonicalization to also canonicalize
the custom header import, although this is unrelated to the above bug.