consider `for_each_cacheless` for HAMT and KAMT

[LesnyRumcajs]

Consider implementing for_each_cacheless, similar to what was done in #2189, to limit memory usage during a single collection iteration.

https://github.com/filecoin-project/ref-fvm/tree/master/ipld/hamt
https://github.com/filecoin-project/ref-fvm/tree/master/ipld/kamt

Finding a decent existing use case to note the real-world performance gain would be great.

[LesnyRumcajs]

@rvagg Do you have some suggestions on an RPC method that takes large amount of memory due to iteration in HAMT? Like Filecoin.StateMarketDeals for AMT?

Or some other existing iteration in the state I could benchmark.

[rvagg]

migrations are where this hurts the most, you could just optimise a migration by doing this in your all-actors iteration

[rvagg]

This beasty here: https://github.com/ChainSafe/forest/blob/905a93e04e6cc6187a421df0d1950bbb124ec43d/src/state_migration/common/state_migration.rs#L107

[LesnyRumcajs]

@hanabi1224 might have some draft for this up for discussions soon.