Cache the _location_without_md5.

[jaraco]

Fixes #2089. Alternate to #2090.

[gotcha]

This relies on the fact that location should be set only once. Is there anything that avoids to set it twice ?
This is why #2090 makes location a property.

[jaraco]

There's nothing that prevents it from being set again, but there's also no reason to think that it will be set again.

[gotcha]

@jaraco While reading the code to answer to your question on #2089, I notice the following:

• hashcmp is mainly slow because of _remove_md5_fragment but it uses a lot of properties
• key and parsed_version are also cached, with two very simple but differing implementations, key uses try except AttributeError; parsed_version uses hasattr

This leads me to the following questions:

• Should the full hashcmp be cached ? I tend to answer yes because it already relies on other cached properties.
• Could it be cached simply, like key or parsed_version ? I also tend to answer positively.
• Which implementation should we choose ? hasattr or try except ? I am not deep enough in Python internals to know which is quicker.
• Is there a reason to use a different implementation of caching for keyand parsed_version ? Should we unify on the quickest ?
• What is the reason to use 3.8 cached_property instead of a simpler caching implementation ?

[jaraco]

I prefer #2108. I'll answer the questions in #2089.