Added docs on fundamental accuracy limits, condition number of the problem

[ahbarnett]

Especially in 1D transforms, especially in float32, we had been getting questions about the large errors that FINUFFT produces. This led to me rewriting and expanding the section on rounding errors, here:
https://finufft.readthedocs.io/en/latest/trouble.html
See under
Mathematical “issues” and error size.

The upshot is that due to the condition number of the mathematical problem, rounding errors (machine precision, not the tol setting) may inevitably be amplified, and that no NUFFT algorithm could avoid this even in principle.

I hope this clarifies some questions we got. Best, Alex