Tests demonstrating inconsistencies in error handling when skip_unavailable=true

[smalyshev]

In relation to the recent skip_unavailable=true change, there's still an inconsistency there. The change in #128163 covers most runtime errors, however it does not cover one code path, namely the one covered by CssPartialErrorsActionListener.onFailure - which handles failures that happen in analyzedPlan . There the code still has this condition:

        if (e instanceof NoClustersToSearchException || ExceptionsHelper.isRemoteUnavailableException(e)) {

which controls the code path where all remotes that we tried when attempting to resolve index failed, and we're about to return an empty result (we're talking about skip_unavailable=true here).
However, if the error that happened on the remote is not a disconnect error but some other error, unlike the runtime branch, skip_unavailable=true would not cover it, and we would still return a failure.

In addition to that, when other indices are present, the same error would instead be hidden by mergedMappings because it is currently only processes unavailability errors.