There's no functionality similar to v2's fromEncoded in v3 yet. While
the encoding does do polyline simplification for loading, I'd suggest
you adapt a similar technique (ala the Ramer-Douglas-Peucker
algorithm) to your current dataset. It'd be better for your servers
and your users if you're not throwing 75k points at the maps api.
While I get the idea of the technique, I won't pretend to fully
understand all of the math behind it. I didn't pay that much attention
in geometry class unfortunately. But I think there's enough code
samples out there to get you where you need to go...