Changing time and dt inside kernel loop to float

[erikvansebille]

As discussed in #2130, using np.timedelta64 and np.datetime64 as core data structured for particles.dt and particles.time brings some major issues:

1. Users will have to convert from timedelta to float inside each kernel, and use the float for multiplying with velocities
2. There is a fundamental issue with uneven timedelta64 dt, see Issue with uneven dt and RK advection #2380

This PR therefore implements a revert to the v3 approach, where time and dt are floats throughout the Kernel. This also allows (later) integration with pint for unit-aware computations in the Kernel

During development, I also came across a bug in RK45 (fixed in 18a4590), which needs a particles.next_dt variable in order to be able to increase dt in the Kernel loop (otherwise, displacements are computed with the old dt, but the time-update is with the new dt).

[VeckoTheGecko]

Only managed to do a partial review today, looks good for the most part - will finish tomorrow. I went ahead and pushed the refactor e479387