Allow ILM's Shrink Action to progress successfully on indices with 1 shard

[bmcconaghy]

When a user specifies the shrink action and the index it is supposed to shrink already has only 1 shard, this error crops up:

    "failed_step": "shrink",
    "phase": "warm",
    "action_time": "1535643967009",
    "rollover_alias": "thing-alias",
    "phase_time": "1535643367314",
    "step_info": "{\"type\":\"illegal_argument_exception\",\"reason\":\"can't shrink an index with only one shard\"}",

ILM should just noop this (or shrink should).

[elasticmachine]

Pinging @elastic/es-core-infra

[talevy]

I do not like the idea of a noop because then the semantics of shrink would not be the same. The shrink should only succeed if it successfully made a new shrunken index. That would mean allowing shrink to operate on an index with one shard, and simply copy without reducing the number of shards (since it cannot).

I've added team-discuss to discuss whether we want to do anything about this. I see there
being three options

1. leave as is. expect an error to occur and be OK with it.
2. make ILM do some roundabout work so that it is as if a shrink occurred (resulting in a shrink-<original_index> index)
3. make the Shrink API allow n-to-n shrinks, where n is the number of shards.

[elasticmachine]

Pinging @elastic/es-core-features

[talevy]

from discussion with the team on Zoom:

We decided it would make sense for ILM shrink
steps to conditionally check whether they should
execute on indices with one shard. If index has
only one shard, the steps will be no-ops and the
policy will continue to run the following actions