Support execution of metrics on a remote host

[matanor]

• A metrics http-based service (under service/metrics), optionally supports token-based authentication.
• A new dockerfile and relevant commands in makefile which exemplify how to containerize a unitxt metrics service.
• The evaluate(..) tool supports usage of remote metrics (only MetricPipeline objects at this point).
• The remote metrics client configuration is loaded from environment variables.

New requirements:

• httpretty - for mocking http servers, in tests

[matanor]

@elronbandel re moving the new pydantic dependency from base.rqr to a new service.rqr file --

The thing is that pydantic.BaseModel is used as a base class for two new classed in metric_utils (here and here). So it is imported in the module level here. If i move the dependency, then this module becomes unusable, right?

If this module was only used within a function, i would have just made a local import. Do you know if there is any similar alternative for the above case?

Discussed an alternative solution suggested by @elronbandel :
We will not use pydantic at all. The MetricRequest and MetricResponse classes will inherit from unitxt DataClass. Then will use to_dict() to convert the request into json (with json.dumps too). In the other direction, will simply construct the MetricResonse from the json response using the default DataClass constructor.

[elronbandel]

We have a mechanism to add and read new environment variables in 'settings_utils.py' I think it's better if you use it. It gives us more control over the settings and allow to review all the settings in one place. It also log new settings changes and have some more nice features.