If we exposed a symbol in an import and used it both unqualified
and qualified, we'd produce an unused warning false positive.
This happened because we were using a single bit flag to determine
whether a value was used qualified or unqualified.
Now that imports can be limited to smaller scopes than the entire module,
unused import warnings need to work like unused def warnings.
This commit moves unused import warnings discovery and reporting from load
to canonicalization where we can track their usage per scope.
This also fixes a longstanding bug where unused exposed names from an import
were not reported if they were only used in a qualified manner.
