log_metric() silently stores NaN, inf, and -inf without raising

Description

mlflow's `_validate_metric()` uses `isinstance(v, numbers.Number)` which returns True for `float('nan')`, `float('inf')`, and `float('-inf')`. No `math.isfinite()` check is performed. All three values are stored silently. Downstream code reading metrics back may receive NaN without any indication the upstream call was semantically invalid. Different backends (SQLAlchemy, file store, REST) handle these values inconsistently.

Reproduction

from mlflow.utils.validation import _validate_metric
import math
_validate_metric('loss', float('nan'),  1000, 0)  # no exception
_validate_metric('loss', float('inf'),  1000, 0)  # no exception
_validate_metric('loss', float('-inf'), 1000, 0)  # no exception
# All three values stored silently in the metrics database