Move benchmarks to benches module with noinline wrappers

[nipunn1313]

This ensures that the benchmarks are in a separate crate and linked
against the smallvec dynamic library rather than being compiled
together.

It also nicely removes the need for the "benchmarks" feature.

---

This change is

[nipunn1313]

Used a #noinline wrapper from benchmarks to get more consistent results.
Before/after shown here.

Looks like the results are a bit worse, but more consistent/realistic since the compiler's inlining during the benchmark itself won't affect the performance.

Nipunns-MBP:rust-smallvec nipunn$ git checkout master ; cargo bench --features benchmarks bench ; git checkout benchcrate ; cargo bench bench
Switched to branch 'master'
Your branch is up-to-date with 'origin/master'.
   Compiling smallvec v0.2.0 (file:///Users/nipunn/src/rust-smallvec)
    Finished release [optimized] target(s) in 2.40 secs
     Running target/release/deps/smallvec-a59689f47ae774cb

running 4 tests
test bench::bench_extend  ... bench:         154 ns/iter (+/- 22)
test bench::bench_insert  ... bench:       1,144 ns/iter (+/- 520)
test bench::bench_push    ... bench:         277 ns/iter (+/- 74)
test bench::bench_pushpop ... bench:         352 ns/iter (+/- 34)

test result: ok. 0 passed; 0 failed; 0 ignored; 4 measured

Switched to branch 'benchcrate'
   Compiling smallvec v0.2.0 (file:///Users/nipunn/src/rust-smallvec)
    Finished release [optimized] target(s) in 2.28 secs
     Running target/release/bench-ef222b4531a86f9a

running 4 tests
test bench_extend  ... bench:         146 ns/iter (+/- 24)
test bench_insert  ... bench:       1,324 ns/iter (+/- 209)
test bench_push    ... bench:         439 ns/iter (+/- 170)
test bench_pushpop ... bench:         360 ns/iter (+/- 218)

test result: ok. 0 passed; 0 failed; 0 ignored; 4 measured

     Running target/release/deps/smallvec-a59689f47ae774cb

running 0 tests

test result: ok. 0 passed; 0 failed; 0 ignored; 0 measured

[nipunn1313]

I verified by disassembling the generated executables that beforehand, insert() was being inlined, but after this change it is not. This accounts for the dip in the bench_insert

[nipunn1313]

Ping on these improvements to the benchmarks.

[jdm]

@bors-servo: r+
Seems reasonable! Thanks!

[bors-servo]

📌 Commit 59f5ea6 has been approved by jdm

[bors-servo]

⚡ Test exempted - status