Override is_available_while_running on DerivedMetric (#5028)

Summary:
Pull Request resolved: #5028

`DerivedMetric.is_available_while_running()` currently inherits the base
`Metric` default of `False`.  This causes `fetch_data_prefer_lookup` to
skip derived metrics for running trials -- it returns cached data (which
is empty) instead of computing from freshly-fetched base data.

Override the classmethod to return `True` so derived metrics always
attempt to compute.  If base data isn't available, `fetch_trial_data`
returns empty data which is handled gracefully downstream.

Reviewed By: lena-kashtelyan

Differential Revision: D96558279

fbshipit-source-id: 5ae71ecf24aa01f7a11a294ab3579e44f092d1fd

Author: ItsMrLin

ax/core/derived_metric.py

@@ -107,6 +107,16 @@ def __init__(
         self._relativize_inputs = relativize_inputs
         self._as_percent = as_percent
 
+    @classmethod
+    def is_available_while_running(cls) -> bool:
+        # Derived metrics should always attempt to compute from base data.
+        # If base data isn't available for running trials, fetch_trial_data
+        # will return empty data, which is handled gracefully.
+        # TODO: Ideally this would dynamically check whether all input metrics
+        # are available while running, but that requires experiment context
+        # that this classmethod doesn't have access to.
+        return True
+
     @property
     def input_metric_names(self) -> list[str]:
         """Names of metrics that this metric depends on."""