Incorrect number of partitions in Comet broadcast operator #243

Describe the bug

This is found during working on #213. See #213 (comment).

Simply, in Spark, broadcast operator is a special operator because it doesn't actually scheduled to execute with correct number of partitions. Spark executes its child operator, collects the result and distributes that. Other operators, when executing, retrieve the broadcasted value.

But in Comet, we simulates this by collecting the child operator's result like Spark, and makes an RDD with each partition containing the broadcasted value. The broadcast operator will be scheduled as normal operator, although it does't contain any execution logic. So we need to assign it a correct number of partitions.

Steps to reproduce

No response

Expected behavior

No response

Additional context

No response