bench: Improve the spectralnorm shootout benchmark

[alexcrichton]

This improves the spectralnorm shootout benchmark through a few vectors after
looking at the leading C implementation:

• The simd-based f64x2 is now used to parallelize a few computations
• RWLock usage has been removed. A custom parallel function was added as a
  form of stack-based fork-join parallelism. I found that the contention on the
  locks was high as well as hindering other optimizations.

This does, however, introduce one unsafe block into the benchmarks, which
previously had none.

In terms of timings, the before and after numbers are:

$ time ./shootout-spectralnorm-before
./shootout-spectralnorm-before  2.07s user 0.71s system 324% cpu 0.857 total
$ time ./shootout-spectralnorm-before 5500
./shootout-spectralnorm-before 5500  11.88s user 1.13s system 459% cpu 2.830 total
$ time ./shootout-spectralnorm-after
./shootout-spectralnorm-after  0.58s user 0.01s system 280% cpu 0.210 tota
$ time ./shootout-spectralnorm-after 5500
./shootout-spectralnorm-after 5500  3.55s user 0.01s system 455% cpu 0.783 total

[rust-highfive]

Warning

• These commits modify unsafe code. Please review it carefully!

[gereeter]

I don't think that this function is safe in general, because f could mutate something it captured. However, it should be possible to encode the correct bound with unboxed closures.

[huonw]

FWIW, I experimented with a similar rewrite that used unboxed closures, and they worked well enough, i.e. no ICEs etc.

[alexcrichton]

Oh dear, you're right! I thought the Sync found was enough, but I think this needs to take an instance of Fn to ensure there's no &mut captures. I've updated with the unboxed closure equivalent of Fn, thanks!

[TeXitoi]

Great! I'd r+ if I can ;)

[kud1ing]

See also #18085