Improve compile time (#1989)

## Issue Addressed

Closes #1264

## Proposed Changes

* Milagro BLS: tweak the feature flags so that Milagro doesn't get compiled if we're using BLST. Profiling showed that it was consuming about 1 minute of CPU time out of 60 minutes of CPU time (real time ~15 mins). A 1.6% saving.
* Reduce monomorphization: compiling for 3 different `EthSpec` types causes a heck of a lot of generic functions to be instantiated (monomorphized). Removing 2 of 3 cuts the LLVM+linking step from around 250 seconds to 180 seconds, a saving of 70 seconds (real time!). This applies only to `make` and not the CI build, because we test with the minimal spec on CI.
* Update `web3` crate to v0.13. This is perhaps the most controversial change, because it requires axing some deposit contract tools from `lcli`. I suspect these tools weren't used much anyway, and could be maintained separately, but I'm also happy to revert this change. However, it does save us a lot of compile time. With #1839, we now have 3 versions of Tokio (and all of Tokio's deps). This change brings us down to 2 versions, but 1 should be achievable once web3 (and reqwest) move to Tokio 0.3.
* Remove `lcli` from the Docker image. It's a dev tool and can be built from the repo if required.

validator_client/Cargo.toml

@@ -32,7 +32,7 @@ slog = { version = "2.5.2", features = ["max_level_trace", "release_max_level_tr
 slog-async = "2.5.0"
 slog-term = "2.6.0"
 tokio = { version = "0.3.2", features = ["time"] }
-futures = { version = "0.3.7", features = ["compat"] }
+futures = "0.3.7"
 dirs = "3.0.1"
 directory = { path = "../common/directory" }
 lockfile = { path = "../common/lockfile" }