fallback: Rewrite buffer from array of bytes to bytes

[methane]

bytearray may be best option for buffer.
But PyPy's struct.unpack is slower on bytearray.
So we use bytes for now.

[jfolz]

Seems like PyPy doesn't like it:

unpacking integers (fallback)    0.211348, 0.154564, 0.154378, 0.156010
unpacking bytes (fallback)       0.563791, 0.541599, 0.537913, 0.541811
unpacking lists (fallback)       0.393923, 0.392007, 0.379326, 0.384622
unpacking dicts (fallback)       0.351910, 0.347979, 0.348440, 0.350482

0.4.7:

unpacking integers (fallback)    0.133905, 0.067842, 0.066790, 0.067156
unpacking bytes (fallback)       0.104518, 0.074977, 0.074882, 0.074800
unpacking lists (fallback)       0.310753, 0.299883, 0.295808, 0.298838
unpacking dicts (fallback)       0.283737, 0.262611, 0.276033, 0.262426

[methane]

bytearray on PyPy is slower than bytes.

master:

unpacking integers (fallback)    0.174160, 0.118780, 0.094318, 0.092160
unpacking bytes (fallback)       0.154044, 0.085751, 0.079935, 0.084918
unpacking lists (fallback)       1.458416, 1.300095, 1.300537, 1.306590
unpacking dicts (fallback)       1.170090, 0.995266, 0.996789, 0.991890

bytearray (5c47807):

unpacking integers (fallback)    0.146611, 0.111243, 0.089572, 0.089555
unpacking bytes (fallback)       0.131598, 0.079429, 0.077857, 0.081479
unpacking lists (fallback)       1.345462, 1.204245, 1.210770, 1.205473
unpacking dicts (fallback)       1.028819, 0.927366, 0.927691, 0.929977

[methane]

FYI, I've filed an issue of bytearray inefficiency to PyPy.
https://bitbucket.org/pypy/pypy/issues/2303/structunpack-is-15x-slower-on-bytearray

[methane]

@jfolz Are you OK to this change?

[jfolz]

@methane Frankly the old buffer code is a mystery to me. This looks much cleaner.

[methane]

On discussion in PyPy issue, Armin pointed out that microbenchmarking on PyPy is too optimized
since we use same bytes for each run.
We need more good benchmark...

[jfolz]

What I gather from that thread is that:

1. Data needs to be different for every run (+ possibly random)
2. Results must be properly used in some way

Though, exactly what "properly used" means is unclear.

[methane]

I've tried another benchmark in my local and I think performance difference between bytes and bytearray is significant.

patch and bench

I merge this pull request as is, and I'll discuss about "slice and unpack" performance in other place.