bpo-41513: More accurate hypot()

[rhettinger]

Draft PR for discussion.

Algorithm combines:

• lossless scaling using a power-of-two scaling factor
• accurate squaring using Veltkamp-Dekker splitting
• accurate summation using a variant of the Neumaier algorithm
• differential correction of the square root calculation

Together, these allow the algorithm to simulate 128-bit quad precision throughout.

On my build, the speed slows 50-60%. I've tested with millions of inputs and not yet found a single 1-ulp error. It seems to always be correctly rounded.

https://bugs.python.org/issue41513