Consider retrying requests when gRPC error detail is "broken pipe"

[pgautier404]

In certain Lambda-based environments we have observed and reliably reproduced a number gRPC errors with STATUS_UNAVAILABLE status code and broken pipe error detail as Lambdas try to reuse cache clients stored in global variables. The error is intermittent, but when it occurs it is always on the first transaction a client attempts after Lambda invocation. The error only occurs when global clients are reused as opposed to being constructed at the time of invocation. If we can verify that this error is guaranteed to be indicative of a complete failure of communication between client and server, we might be fine to always trigger a retry in this situation as idempotency would not be an issue.