Refactor argument-handling to use a struct (#103)

## Summary:
In this commit I refactor the argument-generation logic to move most of
the code out of the template and into the type-generator.  This logic
predates #51, and I didn't think to update it there, but I think it
benefits from similar treatment, for similar reasons.

Specifically, the main change is to treat variables as another struct
type we can generate, rather than handling them inline as a
`map[string]interface{}`.  Users still pass them the same way, but
instead of putting them into a `map[string]interface{}` and JSONifying
that, we generate a struct and put them there.

This turns out to simplify things quite a lot, because we already have a
lot of code to generate types.  Notably, the omitempty code goes from a
dozen lines to basically two, and fixes a bug (#43) in the process,
because now that we have a struct, `json.Marshal` will do our work for
us! (And, once we have syntax for it (#14), we'll be able to handle
field-level omitempty basically for free.)  More importantly, it will
simplify custom marshalers (#38, forthcoming) significantly, since we do
all that logic at the containing-struct level, but will need to apply it
to arguments.

It does require two breaking changes:

1. For folks implementing the `graphql.Client` API (rather than just
   calling `NewClient`): we now pass them variables as an `interface{}`
   rather than a `map[string]interface{}`.  For most callers, including
   Khan/webapp, this is basically a one-line change to the signature of
   their `MakeRequest`, and it should be a lot more future-proof.
2. genqlient's handling of the `omitempty` option has changed to match
   that of `encoding/json`, in particular it now never considers structs
   "empty".  The difference was never intentional (I just didn't realize
   that behavior of `encoding/json`); arguably our behavior was more
   useful but I think that's outweighed by the value of consistency with
   `encoding/json` as well as the simpler and more correct
   implementation (fixing #43 is actually quite nontrivial otherwise).
   Once we have custom unmarshaler support (#38), users will be able to
   map a zero value to JSON null if they wish, which is mostly if not
   entirely equivalent for GraphQL's purposes.

Issue: https://github.com/Khan/genqlient/issues/38
Issue: https://github.com/Khan/genqlient/issues/43

## Test plan:
make check

Author: benjaminjkraft

Reviewers: StevenACoffman, dnerdy, aberkan, jvoll, mahtabsabet, MiguelCastillo

Required Reviewers: 

Approved By: StevenACoffman, dnerdy

Checks: ✅ Test (1.17), ✅ Test (1.16), ✅ Test (1.15), ✅ Test (1.14), ✅ Lint, ✅ Lint, ✅ Test (1.17), ✅ Test (1.16), ✅ Test (1.15), ✅ Test (1.14)

Pull Request URL: https://github.com/Khan/genqlient/pull/103

example/generated.go

@@ -9,6 +9,11 @@ import (
 	"github.com/Khan/genqlient/graphql"
 )
 
+// __getUserInput is used internally by genqlient
+type __getUserInput struct {
+	Login string `json:"Login"`
+}
+
 // getUserResponse is returned by getUser on success.
 type getUserResponse struct {
 	// Lookup a user by login.
@@ -71,12 +76,11 @@ query getViewer {
 func getUser(
 	ctx context.Context,
 	client graphql.Client,
-	login string,
+	Login string,
 ) (*getUserResponse, error) {
-	variables := map[string]interface{}{
-		"Login": login,
+	__input := __getUserInput{
+		Login: Login,
 	}
-
 	var err error
 
 	var retval getUserResponse
@@ -92,7 +96,7 @@ query getUser ($Login: String!) {
 }
 `,
 		&retval,
-		variables,
+		&__input,
 	)
 	return &retval, err
 }