[nix-cache] Prevent checking cache twice per package (#2055)

## Summary

Fixes bug described here #2054

A better (but a bit more involved) solution is to parallelize by output.
Currently we parallelize by meta-output (which includes
`__default_output__`). This has two downsides:

* `__default_output__` can be multiple outputs, in which case we don't
parallelize them at all.
* We still do redundant work when `__default_output__` is multiple
outputs. (once as `__default_output__` and once for each individual
outout that forms part of defaults).

## How was it tested?

internal/devpkg/narinfo_cache.go

@@ -5,6 +5,8 @@ import (
 	"fmt"
 	"io"
 	"net/http"
+	"slices"
+	"strings"
 	"sync"
 	"time"
 
@@ -107,13 +109,30 @@ func (p *Package) fetchNarInfoStatusOnce(output string) (bool, error) {
 	type inCacheFunc func() (bool, error)
 	f, ok := narInfoStatusFnCache.Load(p.Raw)
 	if !ok {
-		key := fmt.Sprintf("%s^%s", p.Raw, output)
 		f = inCacheFunc(sync.OnceValues(func() (bool, error) { return p.fetchNarInfoStatus(output) }))
-		f, _ = narInfoStatusFnCache.LoadOrStore(key, f)
+		f, _ = narInfoStatusFnCache.LoadOrStore(p.keyForOutput(output), f)
 	}
 	return f.(inCacheFunc)()
 }
 
+func (p *Package) keyForOutput(output string) string {
+	if output == useDefaultOutput {
+		sysInfo, err := p.sysInfoIfExists()
+		// let's be super safe to always avoid empty key.
+		if err == nil && sysInfo != nil && len(sysInfo.DefaultOutputs()) > 0 {
+			names := make([]string, len(sysInfo.DefaultOutputs()))
+			for i, o := range sysInfo.DefaultOutputs() {
+				names[i] = o.Name
+			}
+			slices.Sort(names)
+			output = strings.Join(names, ",")
+		}
+	}
+	fmt.Println("output: ", output)
+
+	return fmt.Sprintf("%s^%s", p.Raw, output)
+}
+
 // fetchNarInfoStatus fetches the cache status for the package. It returns
 // true if cache exists, false otherwise.
 // NOTE: This function always performs an HTTP request and should not be called