feature: add support for io operations from filenames

[neutrinoceros]

fix #51
opening as a draft for now, I'll need to add tests

There's room for discussion over the exact implementation for this since no matter how it is handled, it will always add a slight overhead to the existing behaviour, either as an additional conditional check or a try/except.
The latter solution is particularly expensive (performance wise) if an exception is typically raised more than 1% of the time, so I went with the former, which should make both load and dump perform decently in any case, to no significant cost.

A side effect is that in order to achieve this properly, I had to add support to load/dump to and from binary files.

[marzer]

RE performance concerns: given my findings in #42 I'd be surprised it came up on a profiler regardless of which route you go, since the vast majority of CPU time is spent in C++ (toml::parse)

[neutrinoceros]

Nice, this is a relief ! However the patch ended up increasing the complexity of your io functions much more than I anticipated, no hard feelings if you don't want it for this reason :)

[marzer]

Had a quick look at your changes and they look fine to me, but ultimately it's up to @bobfang1992 since the python lib is his project, I'm just the shadowy C++ spectre looming in the shadows :)

One concern I have is the addition of the encoding parameter; note that the TOML spec explicitly states that all valid TOML is UTF-8, so if any non-UTF-8 string is passed through to toml++ it will likely fail with a 'malformed UTF-8' error. Shouldn't be a problem if you can ensure that python handles all the encoding changes internally and the C++ lib never sees anything other than UTF-8.

(worth noting that I don't know how pybind11 handles strings; it's possible that it already ensures they're UTF-8?)

[neutrinoceros]

One concern I have is the addition of the encoding parameter; note that the TOML spec explicitly states that all valid TOML is UTF-8,

Thank you, this is an excellent point that I failed to consider.

Furthermore, quoting the spec (re: binary)

for binary data, it is recommended that you use Base64 or another suitable ASCII or UTF-8 encoding

so It seems pretty clear to me that the encoding param should not be exposed in pytomlpp's api. I'll simplify my patch accordingly.

[neutrinoceros]

I see that the wheels are failing to build because some of the tests that I added use pathlib.Path objects as inputs for load and dump, but the builtin open function only gained support for such object in Python 3.6
It should suffice to convert those paths to strings in the tests, let's see.

[bobfang1992]

Hi I fixed this in #53 can you rebase?

[neutrinoceros]

Done !
To be clear, should I remove the third commit, or do you need to keep tests compatible with Python 3.5 ?

[bobfang1992]

I see no harm keeping them the way they are. I think we can merge this if the CI passes.

[neutrinoceros]

Thank you !