[Feature request] Numerical/memory efficient MC integration

As already discussed (with @dustinvtran) a while ago:
I guess there aren't currently any plans for either

• numerical integration methods (Simpson, quadrature...)
• ("memory efficient") (serialized) MC integration methods

If not, is it okay to create them and add to TFP?

Our use-case: we are currently building a fitting library for high-energy physics on top of TensorFlow (zfit), still in it's early stage. We need a lot of normalization integrals in there.

on the mc integration: the tfp.mc.expectation is already a nice function, but we would like to have a function that:

1. can work serialized (so time vs. memory, being able to average over "infinitely" many samples)
   (2. probably already handles the scaling right (so "integration" instead of "expectation value"))
   But 2. has pretty low priority, 1. would be a need though.

Start by reviewing the existing tfp.mc.expectation API and the requirements described for zfit. Clarify whether the first scope is serialized Monte Carlo integration, numerical integration methods, or both, and define the expected scaling and memory behavior. Done should mean an agreed API and tests demonstrating integration over many samples without retaining them all.