Ability to get callExecutor in ClientInterceptor implementation

Various interceptor implementations need to invoke listener callbacks such as in scenarios:
- returning a [FailingClientCall](https://github.com/grpc/grpc-java/blob/v1.35.0/alts/src/main/java/io/grpc/alts/FailingClientCall.java#L34)
- delaying requests and responses
- retrying requests

However, those listener callbacks maybe called in a wrong thread that its user or the channel provider does not intend to use. I believe many interceptor implementations have been doing this. The callbacks should be running in application thread (executed by callExecutor).

`CallOptions` provides `getExecutor()`, but this may return `null`, and in that case the callExecutor should fallback to the channel executor.

In some cases this would not be a big problem, but not in the case of delay, retry, or there are other interceptors switching application thread/running async tasks involved. The problem could be even more difficult to resolve if an interceptor implementation is to be used as a library for others.

Possible solutions:

- In `ManagedChannelImpl`, we add
```java
if (callOptions.getExecutor() == null) { callOptions.withExecutor(channelExecutor); }
```
But if the interceptor uses a different channel, callOptions.getExecutor() may still be null.

- Add a `getExecutor()` method for `Channel` class. Then Interceptor can call `nextChannel.getExecutor()` as a fallback of `callOptions.getExecutor()`.

Problems to some solutions:
- Executor lifecycle.

Start by reading CallOptions.getExecutor(), ClientInterceptor implementations, and the proposed ManagedChannelImpl and Channel entry points. Compare the fallback options for obtaining the callExecutor, including executor lifecycle concerns, and define how listener callbacks from delayed, retried, or failing calls should consistently run on the intended executor.

索引モデルが issue の本文から書いたものです。