Elasticsearch observer improvements to avoid blocking between workers.

[barkbay]

This PR implements a few improvements on the observer:

• The first synchronous observation implemented in Move first ES cluster state observation out of go routine #5783 is no longer invoked if the Elasticsearch Service is not "ready"
• observer.Start() is invoked outside of any lock-protected section
• An APM span added around observer.observe(...)

I eventually decided to not introduce a custom, low, timeout as suggested in #6078 for the following reasons:

• Moving the first synchronous observation out of the lock-protected section should solve the main bottleneck.
• Adding an additional timeout value means adding a new flag or/and implement an implicit default value.
• If the call to _cluster/health is slow then chances are that it'll also be the case for subsequent API calls, not sure why we would treat this one as a special case.

Fixes #6078

[thbkrkr]

I had forgotten that. It's not great. I opened:

• Create an Elasticsearch client for the observer only if needed #6090

[thbkrkr]

I was about to suggest moving the blocking call outside the lock and you beat me :-)
I checked that it doesn't break the fix made for #5812.

LGTM!