Fix when multiple label names normalize to the same string

[frh-td]

Fixes #774.

I am not a Java programmer, so apologies for any mistakes. From my testing, this works well and fixes the issue.

[SuperQ]

In the issue it says "take the last value", but I think this will only take the first matched value.

But I think this is also fine.

It would be nice to document this somewhere.

[hiSandog]

The duplicate-normalized-label case should probably test both collision detection and deterministic naming. A fixture with two CloudWatch dimensions that normalize to the same Prometheus label name would prove the exporter either rejects the metric with a clear error or applies a stable disambiguation rule. That is important because silently overwriting one label value would produce valid-looking but misleading time series.