rpc/client: check the context before sending message

[minh-bq]

We observe failure in TestEthClient/TxInBlockInterrupted which can be reproduced by

go test -test.v -run=^TestEthClient/TxInBlockInterrupted --count=100

=== RUN   TestEthClient/TxInBlockInterrupted
    ethclient_test.go:399: transaction should be nil
--- FAIL: TestEthClient (0.03s)

The testcase expects to get context cancel error when cancelling context passed to the ethclient call before doing the call. However, the rpc client's send does not guarantee that order. It uses select to wait on sending action and context.Done() simultaneously, so it is possible that the branch of sending action is taken and we get the response without error. This commit adds a non-blocking wait on context.Done() to catch the cancelled context before doing the actually send.

[holiman]

I don't want to fix this if it's only to fix a failing test. One might argue that context cancellation should be prioiitized, thus the change is More Correct.
On the other hand, one might argue that if both cases are instantly available, then one does not need to respect the Cancel -- because the context is also passed to write, so we can hand over the decision to the next part.

Not sure what to think here. I think I lean towards preferring the existing code (which would mean that the test needs to be fixed in a different manner)

[minh-bq]

I agree with that. So the possible solution is to remove the context cancel part in that testcase, just keep this part and rename the testcase

go-ethereum/ethclient/ethclient_test.go

Lines 405 to 408 in e9f59b5

	// Test tx in block not found.
	if _, err := ec.TransactionInBlock(context.Background(), block.Hash(), 20); err != ethereum.NotFound {
	t.Fatal("error should be ethereum.NotFound")
	}

[fjl]

Yes, we should fix the test instead.

[minh-bq]

I've created the fix for the test in #28583