When desugaring inline/local resource contents, the current rules are:

1. If the input config specifies compression: gzip, we pass through the
   compression field unchanged and refuse to compress again.  This is
   useful for local resources that are stored precompressed alongside the
   config.
2. If we decide to compress, we set the compression field to gzip.
3. If we decide not to compress, we do not set the compression field,
   ostensibly so precompressed resources in child configs can inherit the
   compression field from the parent.

But case 3 doesn't really work.  Butane doesn't know that the child
resource is precompressed, so is allowed to compress it again, and only
the compressibility heuristics (which technically are not contractual)
prevent it from doing so and producing invalid output.  (There's no way
to represent nested compression, so the file would be written to the
provisioned system compressed.)  And if the parent resource is not
formally precompressed, but is actually just auto-compressed by Butane,
a change to the parent's contents or to the Butane heuristics could cause
the parent to become uncompressed in the future and break the merged
config.

And case 3 creates a problem for an inline/local child resource overriding
an inline/local parent.  If Butane opts to compress the parent but not
the child, the merged config will have uncompressed contents but inherit
"compression: gzip" from the parent, causing a provisioning failure.  And
if the parent/child contents change over time, this problem may suddenly
appear in the future.

Fix this by changing rule 3.  When desugaring a non-precompressed
inline/local resource, always set the compression field to either "" or
"gzip".  This intuitively makes sense: only we know whether we decided to
compress the resource, so we need to say so.

This involves an API change to MakeDataURL().  We now return a pointer to
the selected compression value, or nil if we're passing through the
compression value from the input (case 1).  In the latter case, the
caller should record a 1:1 translation for the compression field, rather
than attributing its value to the inline/local input field.

Fixes #332.