Random: faster seed hashing

[rfourquet]

When calling seed!(rng, seed), the seed is converted into random bytes, which are then expanded to produce the initialization state for rng. The purpose of hashing is to ensure that MyRNG(1) and MyRNG(2) produce uncorrelated streams.

In practice, this call is implemented as seed!(rng, SeedHasher(seed)), assuming that rng implements seed!(::AbstractRNG) for initialization from another RNG.

Previously, SeedHasher worked in three stages:

1. encode seed into bytes
2. hash these bytes using SHA-2
3. expand the resulting digest with an ad-hoc construction

This approach was functional but relatively slow.
This commit replaces stages 2 and 3 with an algorithm designed specifically for seed generation by M. E. O'Neill, described at: https://www.pcg-random.org/posts/developing-a-seed_seq-alternative.html

The implementation is adapted from O'Neill's seed_seq_fe C++ reference (MIT license). NumPy uses the same algorithm for its SeedSequence.

Here are some numbers:

@btime Xoshiro(1)
@btime Xoshiro($(rand(UInt)))
@btime Xoshiro($(rand(UInt, 4)))
@btime Xoshiro($(rand(UInt, 8)))
s = Random.SeedHasher(); @btime rand($s, UInt)

On master:

  398.755 ns (9 allocations: 448 bytes)
  403.610 ns (9 allocations: 448 bytes)
  486.254 ns (9 allocations: 448 bytes)
  998.182 ns (9 allocations: 448 bytes)
  73.213 ns (0 allocations: 63 bytes)

On PR:

  36.807 ns (3 allocations: 256 bytes)
  54.717 ns (3 allocations: 256 bytes)
  144.917 ns (3 allocations: 256 bytes)
  228.738 ns (3 allocations: 256 bytes)
  2.454 ns (0 allocations: 0 bytes)