Merge `data_location` into `experiment_metadata`

The data location field is a piece of metadata scoped to an experiment,
so one would expect it to be returned by experiment_metadata. But it
is not so: data_location is a separate call. This forces two separate
database queries, which is unnecessary since core_plugin calls them
consecutively at the unique call site. Thus, it’s more than just a
cleanliness issue.

Suggest migrating by first making ExperimentMetadata more malleable:

1. Make ExperimentMetadata.__init__ take kwargs only, and provide
   default values for all fields/arguments.
2. Update data providers to always return a non-None value.

While this loses the ability to distinguish whether the metadata is
present, we already weren’t able to distinguish between whether
individual fields were present. For instance, TensorBoard.dev
experiments that were created prior to the introduction of experiment
names and descriptions still returned metadata with creation time
but with experiment_name="".

Then:

3. Add a new data_location field with default value "".
4. Update data providers to populate the new field, and update
   core_plugin to prefer that field if populated.
5. Remove the data_location method.
6. Remove dead data_location implementations from data providers.

Start at the linked call site in core_plugin.py, then trace ExperimentMetadata and the data provider implementations, including their data_location methods. Confirm how the two consecutive calls are used and update the affected providers and consumers. Done means data_location is available through experiment_metadata and the separate method and dead implementations are removed.