processImport seems unnecessarily slow due to graphql.print

`processImport` : https://github.com/ardatan/graphql-tools/blob/master/packages/import/src/index.ts#L69

This function:

1. parses a document from a file,
2. iterates through each definition
3. prints the definition to a string
4. appends the string definition to a string of all definitions
5. create a source from the string
6. return the source

This results in unnecessarily bad performance as seen in this attached image:

However, this can simply be boiled down to:

1. parse the document from a file
2. append definitions to a new source
3. return the source

Which results in a reduction of processing time:

Code illustration:

```
export function processImport(
filePath: string,
cwd = cwdFactory(),
predefinedImports: Record = {},
visitedFiles: VisitedFilesMap = new Map(),
): DocumentNode {
const set = visitFile(filePath, join(cwd + '/root.graphql'), visitedFiles, predefinedImports);

const document = {
kind: Kind.DOCUMENT,
definitions: [],
};

for (const defs of set.values()) {
for (const def of defs) {
document.definitions.push(def);
}
}

return document;
}
```

As far as I can tell this is functionally the same without adding the overhead of printing and then re-parsing every definition.

Start in packages/import/src/index.ts at processImport, then inspect how visitFile returns definitions and how the current implementation prints and reparses them. Done means preserving the returned DocumentNode behavior while avoiding the unnecessary print-and-parse work described in the issue.

Written by the indexing model from the issue text.