Native Python support for units?

The problem with the existing support is that, for example, [length] + [length] should be of type [length], but if you try to do that with NewType, you just get back float or int (types aren't preserved across binary operations). That's fixable for addition and subtraction relatively easily (define and type hint appropriate dunder methods), but it would be more complicated in the case of multiplication and division, because Python's generics are insufficiently advanced to support e.g. [length] / [time] = [speed]. You could hard-code all of those conversions one at a time, and perhaps generate the code somehow, but a lot of fundamental constants have weird units that wouldn't necessarily have a "standard" interpretation, such as the Boltzmann constant ([length]^2 * [mass] * [time]^-2 * [temperature]^-1), and it would not be fun to hard code those as well.

Ideally, a unit library should allow you to multiply and divide whatever by whatever and just make sense of it, but that would require the ability to put non-type arguments into Python's generics. Then we could express the Boltzmann constant's type as something like Quantity[2, 1, -2, 0, -1, 0, 0], where each number indicates the exponent of a given unit. Right now, Python does not let you do that (to the best of my understanding), because the arguments have to be types, not integers.

However, this also raises more philosophical problems. There are seven base units in the SI system, so one might assume that Quantity should have a fixed arity of seven exponents. However, the SI system doesn't cover many commonly-used units, such as bytes. There are also logarithmic units such as the bel (or decibel) and the cent (used in music theory, not to be confused with various currencies), which are subject to more complicated coherence rules than typical (linear) units. And the radian is officially a derived unit of the form m/m (so that the equation s=rθ) is valid, but that would just simplify to dimensionless unless you special-case it somehow (the steradian has the same problem).