builtin strings.replace_n is nondetermistic when patterns overlap

[srenatus]

I believe this has been noted in #1825, but got lost along the way: since map iteration is nondetermistic, so is the ordering in the replacement array built here. Since in strings.Replacer, "old string comparisons are done in argument order", this has implications on what gets replaced.

AFAICT it only affects overlapping matches.

Expected Behavior

No matter how often I run this,

strings.replace_n({"f": "1", "foo":"2"}, "foo")

I always get the same result.

Actual Behavior

> strings.replace_n({"f": "1", "foo":"2"}, "foo")
"1oo"
> strings.replace_n({"f": "1", "foo":"2"}, "foo")
"1oo"
> strings.replace_n({"f": "1", "foo":"2"}, "foo")
"1oo"
> strings.replace_n({"f": "1", "foo":"2"}, "foo")
"1oo"
> strings.replace_n({"f": "1", "foo":"2"}, "foo")
"1oo"
> strings.replace_n({"f": "1", "foo":"2"}, "foo")
"1oo"
> strings.replace_n({"f": "1", "foo":"2"}, "foo")
"1oo"
> strings.replace_n({"f": "1", "foo":"2"}, "foo")
"2"

Steps to Reproduce the Problem

Run the snippet above many times, note the output.

Additional Info

$ opa version
Version: 0.25.0-dev
Build Commit: 467182247-dirty
Build Timestamp: 2020-10-25T11:50:25Z
Build Hostname: Stephans-MacBook-Pro.local
Go Version: go1.15.2

But I don't think that this has much to do with the OPA version used. It was introduced in v0.15.0.