Improving placement

[fulmicoton]

Given one index, we attempt to spread the load on as few indexer as possible to have better split pruning.

The current logic has some pitfalls:

• Assuming the algorithm ends up spreading the load over two indexers, it does not attempt to do that evenly at all.
• If we have one indexer up, and it receives N pipelines, and more indexers are added. If the N pipeline do not exceed the cpu capacity of the one indexer, the new indexers won't get any of the load.

These are not critical (we are always below the load we target), but they seem to be confusing users.

[dojiong]

What about one pipeline per indexer, and split the indexing pipeline into small tasks to increase the parallelism.

Then the desired_num_pipelines could be used to control the split pruning.

And the indexer could make pipelines from different indexes more balanced on one node(based on things like task units for indexing-pipeline)