Fbpic

This release makes the computation of the laser profiles faster, in particular
in the case when the laser is emitted with the antenna and the profile
thus needs to be computed at every time step.

- When using the laser antenna, the laser profile can now be computed on
GPU, if the profile has the flag `gpu_capable=True`.
(see [473](https://github.com/fbpic/fbpic/pull/473))
- The flattened Gaussian laser was refactored and is now much faster to
compute. (see [486](https://github.com/fbpic/fbpic/pull/486))

This release allows FBPIC to run on GPU with the latest version
of `numba`, by resolving a minor compatibility issue
(see [482](https://github.com/fbpic/fbpic/pull/482)).

It also makes the `ExternalField` faster on GPU (see [470](https://github.com/fbpic/fbpic/pull/470)).

This minor release removes restrictions on the use of recent versions of
numba, when running on GPU.

This release introduces a major change to the treatment of particles close to
the axis (see [347](https://github.com/fbpic/fbpic/pull/347)).
As a result, the code is much more robust when a large amount of
particles simultaneously cross the axis, and concentrate in the very first
cell, in the radial direction.

In particular, this avoids problems in PWFA simulations when particles of the
driver can periodically pinch on the axis. In addition, the details of the
fields at the very tip of the bubble (where sheath electrons cross the axis)
are more realistic. As a result of the new treatment of particles, users may
notice that the charge density deposited on the grid, for a uniform
distribution of particles, appears to have a slight non-uniformity near the
axis. This is a known and expected effect, and can be reduced by increasing
the number of macro-particles in the radial direction (p_nr).

In addition to the above major change, a set of minor changes were introduced:
- JIT functions are now cached when running on CPU, which reduces the
compilation time ([451](https://github.com/fbpic/fbpic/pull/451) and
[445](https://github.com/fbpic/fbpic/pull/445))
- The new release fixes a bug that prevented the code to run on CPU, when
a GPU is available ([454](https://github.com/fbpic/fbpic/pull/454)).

This is minor release of FBPIC, with essentially two improvements:
- It fixes a bug with GPUDirect MPI communications that was introduced in
version 0.16.1, when switching to a more extensive use of `cupy`.
(see [440](https://github.com/fbpic/fbpic/pull/440))
- The compilation for multi-threaded CPU runs is now cached between different
simulations, thereby making to the first step of a simulation much faster.
(see [445](https://github.com/fbpic/fbpic/pull/445))

This release uses `cupy` much more extensively in FBPIC, when running on GPU.
As a result, kernel launch overheads are drastically reduced, and small-size
or mid-size simulations will see a significant speed-up. Another consequence
is that FBPIC now requires Python 3 in order to run on GPU.

See [437](https://github.com/fbpic/fbpic/pull/437) for more details.