Add Bfloat16 Benchmark and Benchmark Suite

[isVoid]

This PR adds bfloat16 kernel benchmarks suite, comparing a raw CUDA kernel runtime and a Numba kernel runtime. It is expected to have high overhead without supporting LTOIR.

The profiling shows slowdown:

                 GOLD: simple_kernel(float *)  PY: cudapy::__main__::kernel[abi:v1,cw51cXTLSUwv1sCUt9Ww0FEw09RRQPKzLTg4gaGKFsG2oMQGEYakJSQB1PQBk0Bynm21OiwU1a0UoLGhDpQE8oxrNQE_3d](Array<float, (int)1, C, mutable, aligned>)
Time (%)                                100.0                                              100.0                                                                                                                             
Total Time (ns)                     1164770.0                                          2753366.0                                                                                                                             
Instances                              1000.0                                             1000.0                                                                                                                             
Avg (ns)                               1164.8                                             2753.4                                                                                                                             
Med (ns)                               1152.0                                             2528.0                                                                                                                             
Min (ns)                               1120.0                                             2495.0                                                                                                                             
Max (ns)                               1504.0                                             8992.0                                                                                                                             
StdDev (ns)                              21.1                                              814.3                                                                                                                             
Perf Ratio (PY / GOLD, %): 
Avg (ns)        236.383929
Med (ns)        219.444444
Min (ns)        222.767857
Max (ns)        597.872340
StdDev (ns)    3859.241706
dtype: float64

Contributes to #12

[isVoid]

Update: with NVIDIA/numba-cuda#48 inplace for Numba-CUDA, we will start to see a very low overhead between raw CUDA kernel performance and Numba CUDA kernel performance:

                 GOLD: simple_kernel(float *)  PY: cudapy::__main__::kernel[abi:v1,cw51cXTLSUwv1sCUt9Ww0FEw09RRQPKiLTj0gIGIFp_2b2oLQFEYYkHSQB1OQAk0Bynm21OizQ1K0UoIGvDpQE8oxrNQE_3d](Array<float, (int)1, C, mutable, aligned>)
Time (%)                                100.0                                              100.0                                                                                                                               
Total Time (ns)                     1136068.0                                          1145038.0                                                                                                                               
Instances                              1000.0                                             1000.0                                                                                                                               
Avg (ns)                               1136.1                                             1145.0                                                                                                                               
Med (ns)                               1121.0                                             1152.0                                                                                                                               
Min (ns)                               1119.0                                             1119.0                                                                                                                               
Max (ns)                               1504.0                                             1536.0                                                                                                                               
StdDev (ns)                              21.6                                               53.1                                                                                                                               
Perf Ratio (PY / GOLD, %): 
Avg (ns)       100.783382
Med (ns)       102.765388
Min (ns)       100.000000
Max (ns)       102.127660
StdDev (ns)    245.833333
dtype: float64

[isVoid]

We should add a readme to document how to use the benchmark suite.

[isVoid]

The work of this PR is still important - but as bfloat16 bindings are introduced into numba-cuda proper, perhaps we need a separate way to introduce these benchmark suites. Pending discussion.