Per transport endpoint stats for performance profiling

All the current stats provided by gRPC appear to be either per-stream or global. Stats about the underlying activity of the transport endpoint implementation (e.g. number of write syscalls, sizes of writes) would be helpful for consumers of gRPC to determine if they're using its facilities efficiently (e.g. comparing the number of messages sent on streams with the number of writes done on those streams, to determine if sufficient concurrent writes are being made to aggregate sends efficiently).