Limit search requests concurrency

[guilload]

Reject search requests with TooManyRequests when too many requests are in flight. This should be done at the root search level.

[fulmicoton]

I think we tend to use the following terminology:

• concurrency limiting: process only n request at a time, and have the others wait in queue.
• load shedding: returning an error right away if two many clients are waiting.

and we usually use them in combination.

Here I think @guilload wants load shedding. (we already have concurrency limiting, at least at the leaf search split level)