websocket: optimize C websocket_mask function

[pjknkda]

This change includes the optimization of websocket_mask function which is written in C extension.

Basically the idea is from the cython implementation of another popular Python web framework - aiohttp. (https://github.com/aio-libs/aiohttp/blob/v1.2.0/aiohttp/_websocket.pyx)

Here is a benchmark result (CPython 3.5.1):

# Original
% python -m timeit \
         -n 100000 \
         -s 'import os; from tornado.speedups import websocket_mask; mask = b"asdf"; data = os.urandom(100 * 1024)' \
         'websocket_mask(mask, data)'
100000 loops, best of 3: 113 usec per loop

# Proposed
% python -m timeit \
         -n 100000 \
         -s 'import os; from tornado.speedups import websocket_mask; mask = b"asdf"; data = os.urandom(100 * 1024)' \
         'websocket_mask(mask, data)'
100000 loops, best of 3: 23.1 usec per loop

The change takes more advantages if the length of data gets longer.

len(data)	Original	Proposed	Ratio
1	0.131	0.129	0.98
10	0.157	0.133	0.85
100	0.253	0.140	0.55
1,000	1.24	0.387	0.31
10,000	11.1	2.36	0.21
100,000	113	22.4	0.20
1,000,000	1110	218	0.20

[ploxiln]

minor style tip, feel free to ignore: this sort of thing is more succinctly written as

    *(uint64_t*)buf = *(uint64_t*)data ^ uint64_mask;

[ploxiln]

another silly little thing which probably has no measurable effect: this can just be if (data_len >= 4) because it can only be true once, saving an extra compare/branch after the first (only) iteration

[pjknkda]

In the case of 32bit system, above 64-bit XOR will not be triggered and it's a fallback for that case.

[ploxiln]

yes, my bad, I see the sizeof() check

[bdarnell]

Thanks!