Pymapd

This release adds the ability to create a `Connection` object by passing a valid `sessionid` instead of passing `user`, `password`, etc. This allows for an external tool to pass a sessionid to pymapd and allow pymapd to continue in the same session as the external tool.

Additionally, a bug was fixed where localized dates threw an error when using `load_table_columnar`.

_Notable PRs_
Update Thrift bindings to 4.6.1 (228)
Detect localized timestamp properly (230)
Ability to connect using sessionid instead of username and password (231)

This release adds preliminary support for taking an existing dashboard and duplicating its functionality, but changing the underlying source table. Users should be advised that this is in a beta state and should test thoroughly to verify that the result is as they expect.

_Notable PRs:_
Dashboard duplication support (154)

Missed an import for setup.py, which made pip install fail.

In this release, a breaking change was made related to how OmniSci version 4.6+ handle dates on the backend. The new behavior is related to storing dates as epoch days instead of seconds. This means the following:

* __for MapD/OmniSci <= 4.6, YOU MUST USE pymapd < 0.11__ (a RuntimeError is thrown if version < 4.6 detected)
* For OmniSci 4.6+, pymapd >= 0.11 is required. No error message will be thrown if using previously released versions of pymapd (i.e. these versions are in the wild, no way to change previously downloaded versions)

Additionally, python 3.5 was dropped, as f-strings are now used within pymapd (a 3.6+ feature). pyarrow 0.13 is also now supported.

_Notable PRs:_

Update build files for pyarrow 0.13 (195)
Remove python 3.5 compat (207)
Swap itertuples for .values (210)
Make DATEs consistent with respect to changes in Omnisci Core (188)
Throw exception if version number <4.6 (220)

This release fixes a library design issue where string columns were previously uploaded as `TEXT` type in OmniSci, when the more desirable behavior is likely `TEXT ENCODING DICT`. Additionally, pandas `categorical` columns will now be uploaded as `TEXT ENCODING DICT`, which previously threw an error about `Unhandled type` error.

Finally, Int8 support was fixed when using the Arrow loader. Additionally, the prior version of the Arrow loader mutated the input data frame in certain situations. This behavior is now remedied by making a copy of the dataframe if a type conversion is needed for upload into OmniSci.

_Notable PRs_
Add dict encoding to all string columns for load (186)
Trim input sql strings (187)
Force categorical to upload as TEXT ENCODED DICT (189)
Remove erroneous int8 cast code and fix mutating df (193)

For 0.9.0, we set lower bounds for pandas, numpy, sqlalchemy and pytest to 1) hopefully speed up conda solving for environments and 2) lower the surface area of acceptable package versions to test against. For users of individual conda environments or virtualenv, this should generally have no adverse affects.

We also removed the ability to specify the format of the Thrift message result on the `Cursor` class (kwarg: `columnar`). This was an internal detail not expected to be needed by the end user; removing it makes the code simpler to maintain.

Finally, we fixed 68, which prevented the ability to get results from tables having columns of arrays.