fix: Load metrics from centralised folder

[ruivieira]

In order to perform evaluations in offline mode, the HF evaluate metrics must be present.
Since metrics are loaded from the current working directory, this could cause conflicts whenever a dataset had the same name as the metric (e.g. glue dataset and glue metric). LMEval would try to load the metric as the dataset.
This issue is more prevalent with unitxt, where there is no way to modify the default loading path from LMEval.
In order to avoid conflicts and still use unitxt in offline mode:

• The metrics were moved to a central location (metrics, avoiding conflicts)
• If using online mode, these local metrics are not used
• If using offline mode, the evaluate load method is patched to use the local metrics (or provided folder)

It is possible to provide a custom metrics folder by supplying the LMEVAL_METRICS_PREFIX environment variable. If not specified the bundled metrics folder is used.

Description

• Move the metrics to a centralised folder
• Modify the evaluate metrics loader to load from supplied path (only in offline mode)

How Has This Been Tested?

Exploratory testing using scenarios in https://github.com/trustyai-explainability/reference/blob/main/lm-eval/LM-EVAL-NEXT.md

Merge criteria:

• The commits are squashed in a cohesive manner and have meaningful messages.
• Testing instructions have been added in the PR body (for PRs involving changes that are not immediately obvious).
• The developer has manually tested the changes and verified that the changes work

[danielezonca]

lgtm