feat(storage-proofs): partial caching for SDR

[dignifiedquire]

• reduces memory usage during precommit_phase1 by 56GiB for 3i2GB sectors, when using max caching
• generate and store parents cache on disk
• mmap cache to use it
• only mmap part of it, to minimize memory usage
• cache is local, not global anymore
• cache can be used by all sector sizes
• enables cache usage by default, given its small overhead (112 KiB at the current configuration)
• increases required fast disk space by 56GiB for 32GiB sectors, this file is stored in /var/tmp/filecoin-parents for now

Based on the ideas in #1135

[cryptonemo]

Looks good. Is the CI error real?

[dignifiedquire]

Looks good. Is the CI error real?

I believe it is a bug in the tests, but will investigate

[porcuquine]

One blocking request to make the cache size a setting.

Otherwise looks good, once CI is fixed. If fixing CI means changing the test, I'd like to see what's going on before we merge.

[porcuquine]

This should go in settings.rs.

[dignifiedquire]

done

[porcuquine]

Thank you. If there are constraints on this value we know need to be honored, please document those too (in settings.rs). Does it need to evenly divide the number of nodes, for example? People should be allowed to shoot themselves in the foot, but we should warn them where the foot is, too.

[deltazxm]

no mem cache anymore?I think this will affect multiple sector sealing speeds.mem cache still necessary

[dignifiedquire]

no mem cache anymore?

there is still a mem cache, it is simply done differently