Add chunk batching

[sorenmulli]

Result of running the timing example shown below.
On GPU, batching gives 2x speedup. On CPU 1.6x. Batched GPU is >5x speedup compared to batched CPU.
Slightly underwhelming numbers but is probably considerably better on GPU for longer texts, haven't tested thoroughly.

➜ CUDA_VISIBLE_DEVICES=0 python scripts/test_timing.py
>>> Profiling device cpu on batch size 1
Initialization time 1.187613
Average time: 1.316037
Std. time: 0.044134
>>> Profiling device cpu on batch size 16
Initialization time 1.461619
Average time: 0.809319
Std. time: 0.026769
>>> Profiling device cpu on batch size 32
Initialization time 1.456767
Average time: 0.791711
Std. time: 0.010899
>>> Profiling device cpu on batch size 64
Initialization time 1.402210
Average time: 0.807623
Std. time: 0.012546
>>> Profiling device cuda on batch size 1
Initialization time 2.387370
Average time: 0.310700
Std. time: 0.000787
>>> Profiling device cuda on batch size 16
Initialization time 1.499094
Average time: 0.150101
Std. time: 0.000458
>>> Profiling device cuda on batch size 32
Initialization time 1.382926
Average time: 0.151230
Std. time: 0.001125
>>> Profiling device cuda on batch size 64
Initialization time 1.308398
Average time: 0.151791
Std. time: 0.015773

[zer0x42]

lgtm