starlark: optimize TestProfile

Previously, the test would compute the first 100,000
elements of the Fibonacci sequence, holding them all in
an array. Since elements may have over 20,000 digits,
this allocates a lot of memory, which, under tight
ulimits of virtual address space, in conjunction with
the mmap int optimization, may cause the process to
have insufficient virtual address space to allocate
the necessary memory.

This change causes it to compute but not retain the
elements of the sequence.

Fixes google/starlark-go#455
diff --git a/starlark/profile_test.go b/starlark/profile_test.go
index 2781833..515d7d4 100644
--- a/starlark/profile_test.go
+++ b/starlark/profile_test.go
@@ -31,10 +31,10 @@
 
 	const src = `
 def fibonacci(n):
-	res = list(range(n))
-	for i in res[2:]:
-		res[i] = res[i-2] + res[i-1]
-	return res
+	x, y = 1, 1
+	for i in range(n):
+		x, y = y, x+y
+	return y
 
 fibonacci(100000)
 `