Adjust codegen-units in optimized builds

[brson]

Would be nice to have this. Note the issues blocking this, listed in a TODO in tikv's Cargo.toml:

• rustc takes more memory to finish release compilation rust-lang/rust#50199
• 35% performance regression in generated code since 1.24 rust-lang/rust#53833

[brson]

Related to #4189 - we probably only want multiple code-gen units in "dev optimized" builds, not true published release builds.

[brson]

Close seems to be accidental.

[brson]

I wrote a script to capture max RSS for rustc

https://github.com/brson/measure-rustc-rss

command: RUSTC_WRAPPER="measure-rustc-rss.sh" make x-build-dev-nopt
toolchain: rustc 1.35.0-nightly (a9da8fc9c 2019-03-04)
platform: Linux

[profile.dev]
opt-level = 0
debug = false
rpath = false
lto = false
debug-assertions = true
codegen-units = 1
panic = "unwind"
incremental = true
overflow-checks = true

debug=false, panic=unwind, opt-level=0, codegen-units=1, incremental=true, lto=false:

tikv rss: 1728404k
tikv-(server|importer|ctl) rss: 3872556k

debug=false, panic=unwind, opt-level=0, codegen-units=16, incremental=true, lto=false:

tikv rss: 1729776k
tikv-(server|importer|ctl) rss: 3964692k

codegen-units on it's own appears to have little impact on memory usage. maybe it's in combination with full LTO that the rss gets bad