Untied RN mode

[pablooliveira]

By setting z = rng::uniform() sample to 0.5; we get an untied round-to-nearest (tie away from zero) mode.
This PR implements such a mode and adds some tests.

During design a compromise arose:

• we can achieve zero-hit performance by using a compile-time template to select between rng::uniform() and 0.5; yet this implies that the rounding mode is set once and for all at the beginning;
• we can allow changing the rounding mode dynamically, which I would like to do in our LLM explorations, but this requires an IF branch.

In the end I decided to implement the IF branch since its overhead cost should be very low, we will pay the branch miss-prediction only the first time after a change of mode. Later the branch should be well predicted and have little overhead.

[yohanchatelain]

Nice PR! I agree with you about the low overhead and to have a dynamic switch is always more convenient to use.