Pymapd

* Unpins Arrow and Pandas dependencies that were held back.

* Clarify GPU installation procedure.

Updates to use cudf 0.15, and bumps a number of dependencies that were held back from cudf 0.14 constraints.

**Note that `cudf 0.15` drops Python 3.6 support and now supports Python 3.7 and 3.8**

Changes:

* Relax RBC dependency constraint to allow newer versions.

* Fix incorrect values returned for `TIMESTAMP` types of millisecond
precision.

* Makes `geopandas` an optional runtime dependency to allow for more reliably solving dependencies in conda environments with existing packages.

* Bump `cudf` to `0.14` to resolve breakage when newer versions of `numba` are used. (see: https://github.com/rapidsai/cudf/pull/5033)

- Update thrift bindings for OmniSciDB 5.3 (325)
- Improved datetime support (324)