Limits: Fetch "vrfdigest" from the client instead of letting the client draw randomness

[Stebalien]

Currently, we:

1. Have an EVM precompile to fetch randomness.
2. Have no real limit on size of the "entropy" we allow.

Honestly, that's fine as long as we charge for the hashing (which we can do). However, the current "extern" APIs pass the entire entropy into lotus and copy it. That's not ok.

Instead, we should:

1. Provide an extern that just returns the VRFDigest (https://github.com/filecoin-project/lotus/blob/1ba2f23efcc99a5ce94057cf2a043ecc3376940a/chain/rand/rand.go#L30) for some given epoch (drand or chain).
2. Do all the heavy lifting in the FVM (so we don't need to copy).

In the future, we should replace these syscalls entirely with syscalls that just get the VRFDigest, computing the actual randomness inside the actors.

[Stebalien]

Alternatively, we could drop the precompile entirely. But this precompile is likely very useful (e.g., because it allows users to check precommit, windowpost, etc. randomness).

[Stebalien]

To be clear:

1. Remove the domain separation tag and entropy from the extern.
2. When we get to https://github.com/filecoin-project/lotus/blob/1ba2f23efcc99a5ce94057cf2a043ecc3376940a/chain/rand/rand.go#L168 or https://github.com/filecoin-project/lotus/blob/1ba2f23efcc99a5ce94057cf2a043ecc3376940a/chain/rand/rand.go#L101, hash the first argument to DrawRandomness (with blake2b) and return the 32byte digest.
3. Perform the actual "draw randomness" operation in the FVM itself.

[Stebalien]

We could also do the future work now and bundle this into the FIP, now that we're tackling this issue anyways.

[Stebalien]

We're currently dropping the precompile, but we should still fix this for M2.2.

So far, @arajasek has done part of this in filecoin-project/filecoin-ffi#348.

[arajasek]

#1842 moves the hashing out of the client into the FVM on master. We'll then want to backport this onto the v2 branch, and land BOTH of those in clients.

I'll open a separate issue to describe the consensus-breaking work of moving the hashing further into the actors themselves.