Add API docs

[juntyr]

Supersedes numpy/numpy-user-dtypes#256

[SwayamInSync]

Structure looks good to me, added following reviews to correct some values, suggested some notes and a bit format fixing

[SwayamInSync]

Suggested change

	:value: :math:`2^{-16382} \cdot (1 - 2^{-112})` or approximately :math:`3.36 \cdot 10^{-4932}`
	:value: :math:`2^{-16382}` or approximately :math:`3.36 \cdot 10^{-4932}`

No multiplier is needed here, what you wrote earlier was the "Largest Subnormal Value" (We don't need to document that)

[juntyr]

I took this from https://sleef.org/quad.xhtml for SLEEF_QUAD_MIN

[SwayamInSync]

Maybe they made error as well, IEEE does not define it like this https://en.wikipedia.org/wiki/Quadruple-precision_floating-point_format

you can confirm this from code too,

import numpy as np
from numpy_quaddtype import QuadPrecision, QuadPrecDType

info = np.finfo(QuadPrecDType())

two = QuadPrecision(2)
smallest_normal_theoretical = two ** QuadPrecision(-16382)


print("Are they equal?", info.smallest_normal == smallest_normal_theoretical) # True

sleef_formula = (two ** QuadPrecision(-16382)) * (QuadPrecision(1) - two ** QuadPrecision(-112))
print("Are they equal?", info.smallest_normal == sleef_formula) # False

[SwayamInSync]

I'll also recommend to run this code on Mac as well :) (I'll run later as I get access to my machine)
My current machine supports __float128 so maybe because of aliasing I am getting true results as per standard.

If they made mistake in their implementation then it'll get caught

[SwayamInSync]

A lot of information present in README is missing here, like source installation, TSan build, documentation build, etc.
So ideally we should add those

[SwayamInSync]

You can maybe selectively copy the sections from there to here

[SwayamInSync]

Suggested change

| `np.dot` | Dot product |

We do not support dot function