fix(android): add support for amerabi-v7a (arm NEON 32bit); fixes #30

[Gioee]

I did some tests with the sqlite3 cli with ".timer on"

Here are the major differences using the vector_full_scan function:

vector_init	32 bit (ms)	64 bit (ms)	% Faster (64)
type=FLOAT16,distance=DOT	4	2	+100%
type=FLOAT16,distance=L1	4	2	+100%
type=FLOAT16/BFLOAT16,distance=L2	12	8	+50%

64 bit is the original implementation, 32 bit is the new implementation to support ARM 32 bit (uses scalar double accumulation)