Remove Sidekiq::Client.default since memoization can break sharding with batches

[jonhyman]

I think this is a fairly subtle bug. Basically, in our code, the only time Sidekiq::Client.push gets called is from batch callbacks. I think that the memoization of the pool is causing issues. This is really subtle, I believe it has happened to 1 out of thousands of batches for me, but here's a conceptual idea of how it could happen.

# Sets the default client, will be fixed to shard 1
Sidekiq::Client.default

# Change my pool to be shard 2 (perhaps in an initializer)
Sidekiq.redis = shard_2_pool

b = Sidekiq::Batch.new
b.on(:success, DoSomething)
b.jobs { MyWorker.perform_async }

Then the batch callback may run on shard 1 if this worker runs the final job, because the batch callbacks use Sidekiq::Client.push which uses the memoized default. This will then fail with "no such batch" because the batch lives on shard 2.

Basically, default is now dangerous because Sidekiq::Client's pool can be changed.

[mperham]

The dreadful side effects of trying to avoid an allocation!