On Lamp the startup sync with bitcoind takes approximatively 2 orders of magnitudes more time using my Esplora instance or bs.info than using a local bitcoind RPC.
I think we should detect by how much block is lightningd far from the tip and cache blocks from Esplora so that the biggest above bottleneck (bandwidth) is compensated.
This is halfly a self-assignment and remainder but if someone wants to work on this feel free (just comment here).