When count is more than half the buffer size, an unbuffered read is likely
to be happening in more than half of the read calls, so it's unlikely to be
worth the extra buffer copy.

Benchmarks show this to be true:

before
2048   7.74  (± 0.46%)  1.24× slower
4096   7.39  (± 0.45%)  1.30× slower
6144   7.52  (± 0.57%)  1.28× slower
8192   9.59  (± 0.79%)       fastest

after
2048   7.73  (± 0.64%)  1.24× slower
4096   8.11  (± 0.76%)  1.18× slower
6144   8.61  (± 0.66%)  1.11× slower
8192    9.6  (± 0.89%)       fastest

Sizes 2048 and 8192 have the same behaviour before and after this commit,
so we see similar speeds. However, we see a 10% speedup for 4096 and 6144
bytes as they call unbuffered_read less often.

Benchmark code:
```
require "benchmark"

macro benchmark(size)
  b.report("{{size}}") do
    # Read a 1 gigabyte file
    in = File.open("1gb", "r")
    buffer = uninitialized UInt8[{{size}}]
    while in.read(buffer.to_slice) > 0
    end
    in.close
  end
end

Benchmark.ips do |b|
  benchmark(2048)
  benchmark(4096)
  benchmark(6144)
  benchmark(8192)
end
```