Adds notes to README

Some notes that may be useful for users (particularly prospective users)
are added to the README, including the discovery by @LukasMosser that
GCC >=4.9 is required.

Closes #18

README.md

@@ -103,5 +103,13 @@ For a full example of using Deepwave for FWI, see [this notebook](https://colab.
 
 It is also possible to perform source inversion with Deepwave. Simply set `requires_grad` to `True` on the source amplitude Tensor.
 
+## Notes
+* Deepwave is only tested with Python 3.6.
+* For a reflective free surface, set the PML width to zero at the surface. For example, in 3D and when the PML width on the other sides is 10 cells, provide the argument `pml_width=[0,10,10,10,10,10]` when creating the propagator if the free surface is at the beginning of the first dimension. The format is [z1, z2, y1, y2, x1, x2], where z1, y1, and x1 are the PML widths at the beginning of the z, y, and x dimensions, and z2, y2, and x2 are the PML widths at the ends of those dimensions.
+* To limit wave simulation to the region covered by the survey (the sources and receivers), provide the `survey_pad` keyword argument when creating the propagator. For example, to use the whole z dimension, but only up to 100m from the first and last source/receiver in the y and x dimensions, use `survey_pad=[None, None, 100, 100, 100, 100]`, where `None` means continue to the edge of the model, and the format is similar to that used for `pml_width`. The default is `None` for every entry.
+* [@LukasMosser](https://github.com/LukasMosser) discovered that GCC 4.9 or above is necessary ([#18](https://github.com/ar4/deepwave/issues/18)).
+* Distributed parallelization over shots is supported, but not within a shot; each shot must run on one node.
+* Currently, the forward source wavefield is saved in memory for use during backpropagation. This means that realistic 3D surveys will probably require more memory than is available. This will be fixed in a future release.
+
 ## Contributing
 Your help to improve Deepwave would be greatly appreciated. If you encounter any difficulty using Deepwave, please report it as a Github Issue so that it can be fixed. If you have feature suggestions or other ideas to make Deepwave better, please also report those as Github Issues. If you want to help with the coding, that would be especially wonderful. The Github Issues contain a list of things that need work. If you see one that you would like to attempt, please ask for it to be assigned to you.