Fixing wrong derivative bug and GPU bug

[unalmis]

Once I mark this ready, I can squash merge.

• Resolves [Bug] Derivatives are wrong #158
• Resolves [Bug] Results are wrong on GPU #161
• Resolves Setting 64 bit floating point precision doesn't affect GPU computation #162
• Tests for computation and derivatives that check correctness of interpolation derivatives. Previously on main, all the tests do is check whether the nufft can do function approximation on uniformly random noise with zero mean. There's basically one free parameter in those tests, so errors are less likely to be detected. At a cursory glance, due to linearity, those tests in main may pass more often than they should.
• Fix modeord bug that yields incorrect derivatives

[unalmis]

From the other open pull request it is clear the developers are aware, but recall that you need to top pin jax <= 0.6.2 .

[TomHilder]

Hi please see #160 as I believe it resolves the issue, which stems from jvp having a hard-coded assumption of the default mode ordering (modeord=0), which you aren't using here. Thanks!

[unalmis]

@TomHilder Glad you found it as well! They should buy us a smoothie or something. I am still getting some issues, I'll share if there's some direction to debug if you are working on this also

[unalmis]

@dfm @TomHilder @lgarrison Any ideas on the GPU bug? Don't be fooled by the close tolerance on some of the tests. There are others where the difference is large enough that is cannot be due to floating point errors.