Decouple PDP chain handlers from synchronous ChainSched execution

[LexLuthr]

PDP watchers registered with ChainSched.AddHandler currently run inline on the ChainNotify consumer path. Several of these handlers perform DB polling, DB transactions, and ETH contract reads. If one stalls, ChainNotify stops draining, which can back up the Lotus RPC output buffer and delay unrelated chain-driven work.

Move PDP handlers to an async, coalesced trigger model:

• ChainSched callback only records the latest head and sends a non-blocking signal.
• A single thread processes the signal.
• Process all watchers from this single thread synchronously to avoid incorrect tipset apply.
• Leave PoRep/window handlers on the existing synchronous path.

This should prevent slow PDP cleanup/delete/settlement/proving scheduling paths from blocking ChainNotify while preserving level-triggered PDP behavior.

[ZenGround0]

Do we have any clear cases where this is stalling out nodes ? If so please add it back to mainnet GA. If this is just defensive that's great but it should be in fast follows.

[ZenGround0]

Oh I see this looks like it is contributing to #1248. Since we are separately addressing that issue in #1255 do you think this can wait until fast follows?