fix: sending a result should be best effort (filecoin-project#42)

The threadpool is used for asynchronous tasks. Due to errors it can
happen that the result is no longer needed. Instead of panicking when
there's no receiving end anymore, only log a message.

Concretely this happens in `bellperson`, where we run the multiexp
operation in parallel on the GPU and the CPU. If there is an error
on the GPU part, we also discard the CPU part. We don't wait for the
CPU operations to be finished, but return early. Hence the receiving
will not be available anymore.

ec-gpu-gen/src/threadpool.rs

@@ -1,7 +1,8 @@
 //! An interface for dealing with the kinds of parallel computations involved.
 use std::env;
 
-use crossbeam_channel::{bounded, Receiver};
+use crossbeam_channel::{bounded, Receiver, SendError};
+use log::trace;
 use once_cell::sync::Lazy;
 use yastl::Pool;
 
@@ -55,7 +56,13 @@ impl Worker {
 
         THREAD_POOL.spawn(move || {
             let res = f();
-            sender.send(res).unwrap();
+            // Best effort. We run it in a separate thread, so the receiver might not exist
+            // anymore, but that's OK. It only means that we are not interested in the result.
+            // A message is logged though, as concurrency issues are hard to debug and this might
+            // help in such cases.
+            if let Err(SendError(_)) = sender.send(res) {
+                trace!("Cannot send result");
+            }
         });
 
         Waiter { receiver }