Cycles: Implement Dwivedi guiding for path-traced subsurface scattering

Cycles has supported path-traced subsurface scattering for a while, but while it's
more accurate than other approaches, the increase in noise makes it an expensive option.

To improve this, this patch implements Dwivedi guiding, a technique that is based on
zero-variance random walk theory from particle physics and helps to produce shorter
random walks with more consistent throughput.

The idea behind this is that in non-white materials, each scattering event inside the
medium reduces the path throughput. Therefore, the darker the material is, the lower the
contribution of paths that travel far from the origin is.
In order to reduce variance, Dwivedi guiding uses modified direction and distance sampling
functions that favor paths which go back towards the medium interface.
By carefully selecting these sampling distributions, variance can be greatly reduced, and
as a neat side effect shorter paths are produced, which speeds up the process.

One limitation of just blindly applying this is that the guiding is derived from the
assumption of a medium that covers an infinite half-space. Therefore, at corners or thin
geometry where this does not hold, the algorithm might lead to fireflies.
To avoid this, the implementation here uses MIS to combine the classic and guided sampling.
Since each of those works on one of the three color channels, the final estimator combines
six sampling techniques. This results in some unintuitive math, but I tried to structure
it in a way that makes some sense.

Another improvement is that in areas where the other side of the mesh is close (e.g. ears),
the algorithm has a chance to switch to guiding towards the other side. This chance is based
on how deep the random walk is inside the object, and once again MIS is applied to the
decision, giving a total of nine techniques.

Combining all this, the noise of path-traced subsurface scattering is reduced significantly.
In my testing with the Rain character model and a simple lighting setup, the path-traced
SSS is now actually less noisy than the Christensen-Burley approximation at same render time
while of course still being significantly more realistic.

Differential Revision: https://developer.blender.org/D9932

Use mmap() IO for reading uncompressed .blends

Instead of submitting tons of tiny IO syscalls, we can speed things up
significantly by `mmap`ing the .blend file into virtual memory and directly
accessing it.

In my local testing, this speeds up loading the Dweebs file with all its
linked files from 19sec to 10sec (on Linux).

As far as I can see, this should be supported on Linux, OSX and BSD.
For Windows, a second code path uses `CreateFileMapping` and
`MapViewOfFile` to achieve the same result.

Reviewed By: mont29, brecht

Differential Revision: https://developer.blender.org/D8246

Fix T84496: Cycles: Tile stealing does not work with NLM

Reviewed By: brecht

Maniphest Tasks: T84496

Differential Revision: https://developer.blender.org/D10066

Fix T82351: Cycles: Tile stealing glitches with adaptive sampling

In my testing this works, but it requires me to remove the min(start_sample...) part in the
adaptive sampling kernel, and I assume there's a reason why it was there?

Reviewed By: brecht

Maniphest Tasks: T82351

Differential Revision: https://developer.blender.org/D9445

Fix unreported: Cycles CLI device override doesn't set peer memory usage flag

Reviewed By: brecht

Differential Revision: https://developer.blender.org/D9929

Cycles: Fix debug compilation after tile stealing commit

Cleanup: clang-format

Cycles: Implement tile stealing to improve CPU+GPU rendering performance

While Cycles already supports using both CPU and GPU at the same time, there
currently is a large problem with it: Since the CPU grabs one tile per thread,
at the end of the render the GPU runs out of new work but the CPU still needs
quite some time to finish its current times.

Having smaller tiles helps somewhat, but especially OpenCL rendering tends to
lose performance with smaller tiles.

Therefore, this commit adds support for tile stealing: When a GPU device runs
out of new tiles, it can signal the CPU to release one of its tiles.
This way, at the end of the render, the GPU quickly finishes the remaining
tiles instead of having to wait for the CPU.

Thanks to AMD for sponsoring this work!

Differential Revision: https://developer.blender.org/D9324

Cleanup: clang-format

Cycles: Add command line option for overriding the compute device

The current way of setting the compute device makes sense for local
use, but for headless rendering it it a massive pain to get Cycles
to use the correct device, usually involving entire Python scripts.

Therefore, this patch adds a simple command-line option to Blender
for specifying the type of device that should be used. If the option
is present, the settings in the user preferences and the scene are
ignored, and instead all devices matching the specified type are used.

Differential Revision: https://developer.blender.org/D9086