ToolkitRegistry() should initialize an empty registry

[j-wags]

@jchodera, the current default constructor for ToolkitRegistry will attempt to initialize OE, Amber, and RDK toolkits. The command to initialize an empty ToolkitRegistry is ToolkitRegistry(toolkit_precedence=[]). Revisiting this months later, this is really unintuitive, and I think we should change it (ToolkitRegistry() should initialize an empty registry). Opening a new issue.

Originally posted by @j-wags in #346 (comment)

[jchodera]

If the API is breaking the "principle of least surprise", then the default behavior should definitely be changed.

I originally thought the least surprising default behavior was "do the right thing automatically unless I tell you not to"---that is, "find all the toolkits that are installed and use those". I expected that, if someone wanted fine-grained control over toolkits, they would initialize with specific toolkits or an empty set and add them one at a time, since this is the minority use case.

But I clearly got the API wrong in my debugging in #346 (comment) :)

[mattwthompson]

I agree that ToolkitRegistry() should give me an instance with no wrappers registered into it. I find this type of behavior quite surprising when I ran into it:

In [1]: from openforcefield.utils.toolkits import *

In [2]: reg = ToolkitRegistry()

In [3]: [reg.register_toolkit(tk) for tk in [OpenEyeToolkitWrapper, RDKitToolkitWrapper, AmberToolsToolkitWrapper]]
Out[3]: [None, None, None]

In [4]: reg.registered_toolkits
Out[4]:
[<openforcefield.utils.toolkits.OpenEyeToolkitWrapper at 0x14071f6d0>,
 <openforcefield.utils.toolkits.RDKitToolkitWrapper at 0x144349550>,
 <openforcefield.utils.toolkits.AmberToolsToolkitWrapper at 0x144349910>,
 <openforcefield.utils.toolkits.BuiltInToolkitWrapper at 0x144349a50>,
 <openforcefield.utils.toolkits.OpenEyeToolkitWrapper at 0x1443758d0>,
 <openforcefield.utils.toolkits.RDKitToolkitWrapper at 0x14527e110>,
 <openforcefield.utils.toolkits.AmberToolsToolkitWrapper at 0x14527e050>]

(this is surprising for two reasons, one being the non-empty default registry and also that .register_toolkit should probably check to see if one is already registered)

I wouldn't expect this to be a hugely problematic API break since most functionality silently uses GLOBAL_TOOLKIT_REGISTRY behind the scenes and initializing a custom ToolkitRegistry() often comes with a populated toolkit_precedence argument. Is there a compelling reason to maintain this behavior that I'm not considering?