fix(events): handle subscribe and get/fill cases separately

[rvagg]

After reviewing #12585 I realised the remaining problem was this:

• subscribe-type calls always wants future reverts but only historical reverts when a tipset is specified
• get-type calls only want reverts when a tipset is specified

But we only have one "excludeReverted" argument to Install(). My first go at this was changing that argument to subscribe and using that to differentiate the behaviour and doing the tipsetKeyCid == cid.Undef check just for prefillFilter. But then I realised that for the non-subscribe calls we only ever install+remove the filter immediately, so the logic is kind of unnecessary if we just introduce a second method to deal with that case.

So now we have a Fill() in addition to an Install(), with the former not installing the filter, it just fills it if there's anything to fill and applies the tipsetKeyCid == cid.Undef directly itself. Then Install() can use Fill() for its historic stuff and then install the filter and ignore the reversion exclusion entirely because you always want those reverts in that case.

• eth_subscribe can't get historic events, so it doesn't touch prefillFilter
• SubscribeActorEventsRaw() can get historic events, but you can also specify a tipset (!) which is kind of silly, but perhaps you want to subscribe to the next tipset, which you have, but there aren't events for it yet .. that should work now, including reversions of it.
• eth_newFilter is the interesting case because it's a hybrid API that can be used a bit like eth_subscribe but for polling, but it can also be used for just getting historic events like eth_getLogs so it needs this dual functionality supplied by Install().

[aarshkshah1992]

@rvagg Just need to fix the lint error.